< prev index next >

src/hotspot/share/opto/graphKit.cpp

Print this page

   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "asm/register.hpp"



  26 #include "ci/ciObjArray.hpp"
  27 #include "ci/ciUtilities.hpp"
  28 #include "classfile/javaClasses.hpp"
  29 #include "compiler/compileLog.hpp"
  30 #include "gc/shared/barrierSet.hpp"
  31 #include "gc/shared/c2/barrierSetC2.hpp"
  32 #include "interpreter/interpreter.hpp"
  33 #include "memory/resourceArea.hpp"

  34 #include "opto/addnode.hpp"

  35 #include "opto/castnode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/graphKit.hpp"
  38 #include "opto/idealKit.hpp"

  39 #include "opto/intrinsicnode.hpp"
  40 #include "opto/locknode.hpp"
  41 #include "opto/machnode.hpp"
  42 #include "opto/memnode.hpp"


  43 #include "opto/opaquenode.hpp"
  44 #include "opto/opcodes.hpp"
  45 #include "opto/parse.hpp"
  46 #include "opto/reachability.hpp"
  47 #include "opto/rootnode.hpp"
  48 #include "opto/runtime.hpp"
  49 #include "opto/subtypenode.hpp"
  50 #include "opto/type.hpp"

  51 #include "runtime/deoptimization.hpp"
  52 #include "runtime/sharedRuntime.hpp"

  53 #include "utilities/bitMap.inline.hpp"
  54 #include "utilities/growableArray.hpp"
  55 #include "utilities/powerOfTwo.hpp"
  56 
  57 //----------------------------GraphKit-----------------------------------------
  58 // Main utility constructor.
  59 GraphKit::GraphKit(JVMState* jvms)
  60   : Phase(Phase::Parser),
  61     _env(C->env()),
  62     _gvn(*C->initial_gvn()),
  63     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  64 {

  65   _exceptions = jvms->map()->next_exception();
  66   if (_exceptions != nullptr)  jvms->map()->set_next_exception(nullptr);
  67   set_jvms(jvms);







  68 }
  69 
  70 // Private constructor for parser.
  71 GraphKit::GraphKit()
  72   : Phase(Phase::Parser),
  73     _env(C->env()),
  74     _gvn(*C->initial_gvn()),
  75     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  76 {
  77   _exceptions = nullptr;
  78   set_map(nullptr);
  79   DEBUG_ONLY(_sp = -99);
  80   DEBUG_ONLY(set_bci(-99));
  81 }
  82 
  83 



















  84 
  85 //---------------------------clean_stack---------------------------------------
  86 // Clear away rubbish from the stack area of the JVM state.
  87 // This destroys any arguments that may be waiting on the stack.
  88 void GraphKit::clean_stack(int from_sp) {
  89   SafePointNode* map      = this->map();
  90   JVMState*      jvms     = this->jvms();
  91   int            stk_size = jvms->stk_size();
  92   int            stkoff   = jvms->stkoff();
  93   Node*          top      = this->top();
  94   for (int i = from_sp; i < stk_size; i++) {
  95     if (map->in(stkoff + i) != top) {
  96       map->set_req(stkoff + i, top);
  97     }
  98   }
  99 }
 100 
 101 
 102 //--------------------------------sync_jvms-----------------------------------
 103 // Make sure our current jvms agrees with our parse state.

 902         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 903           tty->print_cr("Zombie local %d: ", local);
 904           jvms->dump();
 905         }
 906         return false;
 907       }
 908     }
 909   }
 910   return true;
 911 }
 912 
 913 #endif //ASSERT
 914 
 915 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
 916 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 917   ciMethod* cur_method = jvms->method();
 918   int       cur_bci   = jvms->bci();
 919   if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
 920     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 921     return Interpreter::bytecode_should_reexecute(code) ||
 922            (is_anewarray && code == Bytecodes::_multianewarray);
 923     // Reexecute _multianewarray bytecode which was replaced with
 924     // sequence of [a]newarray. See Parse::do_multianewarray().
 925     //
 926     // Note: interpreter should not have it set since this optimization
 927     // is limited by dimensions and guarded by flag so in some cases
 928     // multianewarray() runtime calls will be generated and
 929     // the bytecode should not be reexecutes (stack will not be reset).
 930   } else {
 931     return false;
 932   }
 933 }
 934 
 935 // Helper function for adding JVMState and debug information to node
 936 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 937   // Add the safepoint edges to the call (or other safepoint).
 938 
 939   // Make sure dead locals are set to top.  This
 940   // should help register allocation time and cut down on the size
 941   // of the deoptimization information.
 942   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");

1018     uint p = debug_start;  // walks forward in [debug_start, debug_end)
1019     uint j, k, l;
1020     SafePointNode* in_map = in_jvms->map();
1021     out_jvms->set_map(call);
1022 
1023     if (can_prune_locals) {
1024       assert(in_jvms->method() == out_jvms->method(), "sanity");
1025       // If the current throw can reach an exception handler in this JVMS,
1026       // then we must keep everything live that can reach that handler.
1027       // As a quick and dirty approximation, we look for any handlers at all.
1028       if (in_jvms->method()->has_exception_handlers()) {
1029         can_prune_locals = false;
1030       }
1031     }
1032 
1033     // Add the Locals
1034     k = in_jvms->locoff();
1035     l = in_jvms->loc_size();
1036     out_jvms->set_locoff(p);
1037     if (!can_prune_locals) {
1038       for (j = 0; j < l; j++)
1039         call->set_req(p++, in_map->in(k+j));

1040     } else {
1041       p += l;  // already set to top above by add_req_batch
1042     }
1043 
1044     // Add the Expression Stack
1045     k = in_jvms->stkoff();
1046     l = in_jvms->sp();
1047     out_jvms->set_stkoff(p);
1048     if (!can_prune_locals) {
1049       for (j = 0; j < l; j++)
1050         call->set_req(p++, in_map->in(k+j));

1051     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1052       // Divide stack into {S0,...,S1}, where S0 is set to top.
1053       uint s1 = stack_slots_not_pruned;
1054       stack_slots_not_pruned = 0;  // for next iteration
1055       if (s1 > l)  s1 = l;
1056       uint s0 = l - s1;
1057       p += s0;  // skip the tops preinstalled by add_req_batch
1058       for (j = s0; j < l; j++)
1059         call->set_req(p++, in_map->in(k+j));
1060     } else {
1061       p += l;  // already set to top above by add_req_batch
1062     }
1063 
1064     // Add the Monitors
1065     k = in_jvms->monoff();
1066     l = in_jvms->mon_size();
1067     out_jvms->set_monoff(p);
1068     for (j = 0; j < l; j++)
1069       call->set_req(p++, in_map->in(k+j));
1070 

1258   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1259   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1260   return _gvn.transform( new AndLNode(conv, mask) );
1261 }
1262 
1263 Node* GraphKit::ConvL2I(Node* offset) {
1264   // short-circuit a common case
1265   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1266   if (offset_con != (jlong)Type::OffsetBot) {
1267     return intcon((int) offset_con);
1268   }
1269   return _gvn.transform( new ConvL2INode(offset));
1270 }
1271 
1272 //-------------------------load_object_klass-----------------------------------
1273 Node* GraphKit::load_object_klass(Node* obj) {
1274   // Special-case a fresh allocation to avoid building nodes:
1275   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1276   if (akls != nullptr)  return akls;
1277   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1278   return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1279 }
1280 
1281 //-------------------------load_array_length-----------------------------------
1282 Node* GraphKit::load_array_length(Node* array) {
1283   // Special-case a fresh allocation to avoid building nodes:
1284   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1285   Node *alen;
1286   if (alloc == nullptr) {
1287     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1288     alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1289   } else {
1290     alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1291   }
1292   return alen;
1293 }
1294 
1295 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1296                                    const TypeOopPtr* oop_type,
1297                                    bool replace_length_in_map) {
1298   Node* length = alloc->Ideal_length();

1307         replace_in_map(length, ccast);
1308       }
1309       return ccast;
1310     }
1311   }
1312   return length;
1313 }
1314 
1315 //------------------------------do_null_check----------------------------------
1316 // Helper function to do a null pointer check.  Returned value is
1317 // the incoming address with null casted away.  You are allowed to use the
1318 // not-null value only if you are control dependent on the test.
1319 #ifndef PRODUCT
1320 extern uint explicit_null_checks_inserted,
1321             explicit_null_checks_elided;
1322 #endif
1323 Node* GraphKit::null_check_common(Node* value, BasicType type,
1324                                   // optional arguments for variations:
1325                                   bool assert_null,
1326                                   Node* *null_control,
1327                                   bool speculative) {

1328   assert(!assert_null || null_control == nullptr, "not both at once");
1329   if (stopped())  return top();
1330   NOT_PRODUCT(explicit_null_checks_inserted++);
1331 



















1332   // Construct null check
1333   Node *chk = nullptr;
1334   switch(type) {
1335     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1336     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1337     case T_ARRAY  : // fall through
1338       type = T_OBJECT;  // simplify further tests
1339     case T_OBJECT : {
1340       const Type *t = _gvn.type( value );
1341 
1342       const TypeOopPtr* tp = t->isa_oopptr();
1343       if (tp != nullptr && !tp->is_loaded()
1344           // Only for do_null_check, not any of its siblings:
1345           && !assert_null && null_control == nullptr) {
1346         // Usually, any field access or invocation on an unloaded oop type
1347         // will simply fail to link, since the statically linked class is
1348         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1349         // the static class is loaded but the sharper oop type is not.
1350         // Rather than checking for this obscure case in lots of places,
1351         // we simply observe that a null check on an unloaded class

1415         }
1416         Node *oldcontrol = control();
1417         set_control(cfg);
1418         Node *res = cast_not_null(value);
1419         set_control(oldcontrol);
1420         NOT_PRODUCT(explicit_null_checks_elided++);
1421         return res;
1422       }
1423       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1424       if (cfg == nullptr)  break;  // Quit at region nodes
1425       depth++;
1426     }
1427   }
1428 
1429   //-----------
1430   // Branch to failure if null
1431   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1432   Deoptimization::DeoptReason reason;
1433   if (assert_null) {
1434     reason = Deoptimization::reason_null_assert(speculative);
1435   } else if (type == T_OBJECT) {
1436     reason = Deoptimization::reason_null_check(speculative);
1437   } else {
1438     reason = Deoptimization::Reason_div0_check;
1439   }
1440   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1441   // ciMethodData::has_trap_at will return a conservative -1 if any
1442   // must-be-null assertion has failed.  This could cause performance
1443   // problems for a method after its first do_null_assert failure.
1444   // Consider using 'Reason_class_check' instead?
1445 
1446   // To cause an implicit null check, we set the not-null probability
1447   // to the maximum (PROB_MAX).  For an explicit check the probability
1448   // is set to a smaller value.
1449   if (null_control != nullptr || too_many_traps(reason)) {
1450     // probability is less likely
1451     ok_prob =  PROB_LIKELY_MAG(3);
1452   } else if (!assert_null &&
1453              (ImplicitNullCheckThreshold > 0) &&
1454              method() != nullptr &&
1455              (method()->method_data()->trap_count(reason)

1489   }
1490 
1491   if (assert_null) {
1492     // Cast obj to null on this path.
1493     replace_in_map(value, zerocon(type));
1494     return zerocon(type);
1495   }
1496 
1497   // Cast obj to not-null on this path, if there is no null_control.
1498   // (If there is a null_control, a non-null value may come back to haunt us.)
1499   if (type == T_OBJECT) {
1500     Node* cast = cast_not_null(value, false);
1501     if (null_control == nullptr || (*null_control) == top())
1502       replace_in_map(value, cast);
1503     value = cast;
1504   }
1505 
1506   return value;
1507 }
1508 
1509 
1510 //------------------------------cast_not_null----------------------------------
1511 // Cast obj to not-null on this path
1512 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1513   const Type *t = _gvn.type(obj);
1514   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1515   // Object is already not-null?
1516   if( t == t_not_null ) return obj;

1517 
1518   Node* cast = new CastPPNode(control(), obj,t_not_null);
1519   cast = _gvn.transform( cast );
1520 
1521   // Scan for instances of 'obj' in the current JVM mapping.
1522   // These instances are known to be not-null after the test.
1523   if (do_replace_in_map)
1524     replace_in_map(obj, cast);
1525 
1526   return cast;                  // Return casted value
1527 }
1528 
1529 // Sometimes in intrinsics, we implicitly know an object is not null
1530 // (there's no actual null check) so we can cast it to not null. In
1531 // the course of optimizations, the input to the cast can become null.
1532 // In that case that data path will die and we need the control path
1533 // to become dead as well to keep the graph consistent. So we have to
1534 // add a check for null for which one branch can't be taken. It uses
1535 // an OpaqueConstantBool node that will cause the check to be removed after loop
1536 // opts so the test goes away and the compiled code doesn't execute a
1537 // useless check.
1538 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1539   if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1540     return value;
1541   }
1542   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1543   Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1544   Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1545   IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1546   _gvn.set_type(iff, iff->Value(&_gvn));

1585 //=============================================================================
1586 //--------------------------------memory---------------------------------------
1587 Node* GraphKit::memory(uint alias_idx) {
1588   MergeMemNode* mem = merged_memory();
1589   Node* p = mem->memory_at(alias_idx);
1590   assert(p != mem->empty_memory(), "empty");
1591   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1592   return p;
1593 }
1594 
1595 //-----------------------------reset_memory------------------------------------
1596 Node* GraphKit::reset_memory() {
1597   Node* mem = map()->memory();
1598   // do not use this node for any more parsing!
1599   DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1600   return _gvn.transform( mem );
1601 }
1602 
1603 //------------------------------set_all_memory---------------------------------
1604 void GraphKit::set_all_memory(Node* newmem) {
1605   Node* mergemem = MergeMemNode::make(newmem);
1606   gvn().set_type_bottom(mergemem);











1607   map()->set_memory(mergemem);
1608 }
1609 
1610 //------------------------------set_all_memory_call----------------------------
1611 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1612   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1613   set_all_memory(newmem);
1614 }
1615 
1616 //=============================================================================
1617 //
1618 // parser factory methods for MemNodes
1619 //
1620 // These are layered on top of the factory methods in LoadNode and StoreNode,
1621 // and integrate with the parser's memory state and _gvn engine.
1622 //
1623 
1624 // factory methods in "int adr_idx"
1625 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1626                           MemNode::MemOrd mo,
1627                           LoadNode::ControlDependency control_dependency,
1628                           bool require_atomic_access,
1629                           bool unaligned,
1630                           bool mismatched,
1631                           bool unsafe,
1632                           uint8_t barrier_data) {
1633   int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1634   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1635   const TypePtr* adr_type = nullptr; // debug-mode-only argument
1636   DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1637   Node* mem = memory(adr_idx);
1638   Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1639   ld = _gvn.transform(ld);

1640   if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1641     // Improve graph before escape analysis and boxing elimination.
1642     record_for_igvn(ld);
1643     if (ld->is_DecodeN()) {
1644       // Also record the actual load (LoadN) in case ld is DecodeN. In some
1645       // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1646       // a Phi). Recording such cases is still perfectly sound, but may be
1647       // unnecessary and result in some minor IGVN overhead.
1648       record_for_igvn(ld->in(1));
1649     }
1650   }
1651   return ld;
1652 }
1653 
1654 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1655                                 MemNode::MemOrd mo,
1656                                 bool require_atomic_access,
1657                                 bool unaligned,
1658                                 bool mismatched,
1659                                 bool unsafe,

1673   if (unsafe) {
1674     st->as_Store()->set_unsafe_access();
1675   }
1676   st->as_Store()->set_barrier_data(barrier_data);
1677   st = _gvn.transform(st);
1678   set_memory(st, adr_idx);
1679   // Back-to-back stores can only remove intermediate store with DU info
1680   // so push on worklist for optimizer.
1681   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1682     record_for_igvn(st);
1683 
1684   return st;
1685 }
1686 
1687 Node* GraphKit::access_store_at(Node* obj,
1688                                 Node* adr,
1689                                 const TypePtr* adr_type,
1690                                 Node* val,
1691                                 const Type* val_type,
1692                                 BasicType bt,
1693                                 DecoratorSet decorators) {


1694   // Transformation of a value which could be null pointer (CastPP #null)
1695   // could be delayed during Parse (for example, in adjust_map_after_if()).
1696   // Execute transformation here to avoid barrier generation in such case.
1697   if (_gvn.type(val) == TypePtr::NULL_PTR) {
1698     val = _gvn.makecon(TypePtr::NULL_PTR);
1699   }
1700 
1701   if (stopped()) {
1702     return top(); // Dead path ?
1703   }
1704 
1705   assert(val != nullptr, "not dead path");







1706 
1707   C2AccessValuePtr addr(adr, adr_type);
1708   C2AccessValue value(val, val_type);
1709   C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1710   if (access.is_raw()) {
1711     return _barrier_set->BarrierSetC2::store_at(access, value);
1712   } else {
1713     return _barrier_set->store_at(access, value);
1714   }
1715 }
1716 
1717 Node* GraphKit::access_load_at(Node* obj,   // containing obj
1718                                Node* adr,   // actual address to store val at
1719                                const TypePtr* adr_type,
1720                                const Type* val_type,
1721                                BasicType bt,
1722                                DecoratorSet decorators) {

1723   if (stopped()) {
1724     return top(); // Dead path ?
1725   }
1726 
1727   SavedState old_state(this);
1728   C2AccessValuePtr addr(adr, adr_type);
1729   C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1730   Node* load;
1731   if (access.is_raw()) {
1732     load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1733   } else {
1734     load = _barrier_set->load_at(access, val_type);
1735   }
1736 
1737   // Restore the previous state only if the load got folded to a constant
1738   // and we can discard any barriers that might have been added.
1739   if (load == nullptr || !load->is_Con()) {
1740     old_state.discard();
1741   }
1742   return load;
1743 }
1744 
1745 Node* GraphKit::access_load(Node* adr,   // actual address to load val at
1746                             const Type* val_type,
1747                             BasicType bt,
1748                             DecoratorSet decorators) {
1749   if (stopped()) {

1831                                      Node* new_val,
1832                                      const Type* value_type,
1833                                      BasicType bt,
1834                                      DecoratorSet decorators) {
1835   C2AccessValuePtr addr(adr, adr_type);
1836   C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1837   if (access.is_raw()) {
1838     return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1839   } else {
1840     return _barrier_set->atomic_add_at(access, new_val, value_type);
1841   }
1842 }
1843 
1844 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1845   return _barrier_set->clone(this, src, dst, size, is_array);
1846 }
1847 
1848 //-------------------------array_element_address-------------------------
1849 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1850                                       const TypeInt* sizetype, Node* ctrl) {
1851   uint shift  = exact_log2(type2aelembytes(elembt));
1852   uint header = arrayOopDesc::base_offset_in_bytes(elembt);













1853 
1854   // short-circuit a common case (saves lots of confusing waste motion)
1855   jint idx_con = find_int_con(idx, -1);
1856   if (idx_con >= 0) {
1857     intptr_t offset = header + ((intptr_t)idx_con << shift);
1858     return basic_plus_adr(ary, offset);
1859   }
1860 
1861   // must be correct type for alignment purposes
1862   Node* base  = basic_plus_adr(ary, header);
1863   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1864   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1865   return basic_plus_adr(ary, base, scale);
1866 }
1867 

































1868 //-------------------------load_array_element-------------------------
1869 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1870   const Type* elemtype = arytype->elem();
1871   BasicType elembt = elemtype->array_element_basic_type();
1872   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1873   if (elembt == T_NARROWOOP) {
1874     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1875   }
1876   Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1877                             IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1878   return ld;
1879 }
1880 
1881 //-------------------------set_arguments_for_java_call-------------------------
1882 // Arguments (pre-popped from the stack) are taken from the JVMS.
1883 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1884   // Add the call arguments:
1885   uint nargs = call->method()->arg_size();
1886   for (uint i = 0; i < nargs; i++) {
1887     Node* arg = argument(i);
1888     call->init_req(i + TypeFunc::Parms, arg);














































1889   }
1890 }
1891 
1892 //---------------------------set_edges_for_java_call---------------------------
1893 // Connect a newly created call into the current JVMS.
1894 // A return value node (if any) is returned from set_edges_for_java_call.
1895 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1896 
1897   // Add the predefined inputs:
1898   call->init_req( TypeFunc::Control, control() );
1899   call->init_req( TypeFunc::I_O    , i_o() );
1900   call->init_req( TypeFunc::Memory , reset_memory() );
1901   call->init_req( TypeFunc::FramePtr, frameptr() );
1902   call->init_req( TypeFunc::ReturnAdr, top() );
1903 
1904   add_safepoint_edges(call, must_throw);
1905 
1906   Node* xcall = _gvn.transform(call);
1907 
1908   if (xcall == top()) {
1909     set_control(top());
1910     return;
1911   }
1912   assert(xcall == call, "call identity is stable");
1913 
1914   // Re-use the current map to produce the result.
1915 
1916   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1917   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1918   set_all_memory_call(xcall, separate_io_proj);
1919 
1920   //return xcall;   // no need, caller already has it
1921 }
1922 
1923 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1924   if (stopped())  return top();  // maybe the call folded up?
1925 
1926   // Capture the return value, if any.
1927   Node* ret;
1928   if (call->method() == nullptr ||
1929       call->method()->return_type()->basic_type() == T_VOID)
1930         ret = top();
1931   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1932 
1933   // Note:  Since any out-of-line call can produce an exception,
1934   // we always insert an I_O projection from the call into the result.
1935 
1936   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1937 
1938   if (separate_io_proj) {
1939     // The caller requested separate projections be used by the fall
1940     // through and exceptional paths, so replace the projections for
1941     // the fall through path.
1942     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1943     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1944   }




















































































1945   return ret;
1946 }
1947 
1948 //--------------------set_predefined_input_for_runtime_call--------------------
1949 // Reading and setting the memory state is way conservative here.
1950 // The real problem is that I am not doing real Type analysis on memory,
1951 // so I cannot distinguish card mark stores from other stores.  Across a GC
1952 // point the Store Barrier and the card mark memory has to agree.  I cannot
1953 // have a card mark store and its barrier split across the GC point from
1954 // either above or below.  Here I get that to happen by reading ALL of memory.
1955 // A better answer would be to separate out card marks from other memory.
1956 // For now, return the input memory state, so that it can be reused
1957 // after the call, if this call has restricted memory effects.
1958 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1959   // Set fixed predefined input arguments
1960   call->init_req(TypeFunc::Control, control());
1961   call->init_req(TypeFunc::I_O, top()); // does no i/o
1962   call->init_req(TypeFunc::ReturnAdr, top());
1963   if (call->is_CallLeafPure()) {
1964     call->init_req(TypeFunc::Memory, top());

2026     if (use->is_MergeMem()) {
2027       wl.push(use);
2028     }
2029   }
2030 }
2031 
2032 // Replace the call with the current state of the kit.
2033 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2034   JVMState* ejvms = nullptr;
2035   if (has_exceptions()) {
2036     ejvms = transfer_exceptions_into_jvms();
2037   }
2038 
2039   ReplacedNodes replaced_nodes = map()->replaced_nodes();
2040   ReplacedNodes replaced_nodes_exception;
2041   Node* ex_ctl = top();
2042 
2043   SafePointNode* final_state = stop();
2044 
2045   // Find all the needed outputs of this call
2046   CallProjections callprojs;
2047   call->extract_projections(&callprojs, true, do_asserts);
2048 
2049   Unique_Node_List wl;
2050   Node* init_mem = call->in(TypeFunc::Memory);
2051   Node* final_mem = final_state->in(TypeFunc::Memory);
2052   Node* final_ctl = final_state->in(TypeFunc::Control);
2053   Node* final_io = final_state->in(TypeFunc::I_O);
2054 
2055   // Replace all the old call edges with the edges from the inlining result
2056   if (callprojs.fallthrough_catchproj != nullptr) {
2057     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
2058   }
2059   if (callprojs.fallthrough_memproj != nullptr) {
2060     if (final_mem->is_MergeMem()) {
2061       // Parser's exits MergeMem was not transformed but may be optimized
2062       final_mem = _gvn.transform(final_mem);
2063     }
2064     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
2065     add_mergemem_users_to_worklist(wl, final_mem);
2066   }
2067   if (callprojs.fallthrough_ioproj != nullptr) {
2068     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
2069   }
2070 
2071   // Replace the result with the new result if it exists and is used
2072   if (callprojs.resproj != nullptr && result != nullptr) {
2073     C->gvn_replace_by(callprojs.resproj, result);



















2074   }
2075 
2076   if (ejvms == nullptr) {
2077     // No exception edges to simply kill off those paths
2078     if (callprojs.catchall_catchproj != nullptr) {
2079       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
2080     }
2081     if (callprojs.catchall_memproj != nullptr) {
2082       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
2083     }
2084     if (callprojs.catchall_ioproj != nullptr) {
2085       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
2086     }
2087     // Replace the old exception object with top
2088     if (callprojs.exobj != nullptr) {
2089       C->gvn_replace_by(callprojs.exobj, C->top());
2090     }
2091   } else {
2092     GraphKit ekit(ejvms);
2093 
2094     // Load my combined exception state into the kit, with all phis transformed:
2095     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2096     replaced_nodes_exception = ex_map->replaced_nodes();
2097 
2098     Node* ex_oop = ekit.use_exception_state(ex_map);
2099 
2100     if (callprojs.catchall_catchproj != nullptr) {
2101       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2102       ex_ctl = ekit.control();
2103     }
2104     if (callprojs.catchall_memproj != nullptr) {
2105       Node* ex_mem = ekit.reset_memory();
2106       C->gvn_replace_by(callprojs.catchall_memproj,   ex_mem);
2107       add_mergemem_users_to_worklist(wl, ex_mem);
2108     }
2109     if (callprojs.catchall_ioproj != nullptr) {
2110       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
2111     }
2112 
2113     // Replace the old exception object with the newly created one
2114     if (callprojs.exobj != nullptr) {
2115       C->gvn_replace_by(callprojs.exobj, ex_oop);
2116     }
2117   }
2118 
2119   // Disconnect the call from the graph
2120   call->disconnect_inputs(C);
2121   C->gvn_replace_by(call, C->top());
2122 
2123   // Clean up any MergeMems that feed other MergeMems since the
2124   // optimizer doesn't like that.
2125   while (wl.size() > 0) {
2126     _gvn.transform(wl.pop());
2127   }
2128 
2129   if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2130     replaced_nodes.apply(C, final_ctl);
2131   }
2132   if (!ex_ctl->is_top() && do_replaced_nodes) {
2133     replaced_nodes_exception.apply(C, ex_ctl);
2134   }
2135 }
2136 
2137 
2138 //------------------------------increment_counter------------------------------
2139 // for statistics: increment a VM counter by 1
2140 
2141 void GraphKit::increment_counter(address counter_addr) {
2142   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2143   increment_counter(adr1);
2144 }
2145 
2146 void GraphKit::increment_counter(Node* counter_addr) {
2147   Node* ctrl = control();
2148   Node* cnt  = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2149   Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2150   store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2151 }
2152 
2153 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2154   Node* halt = new HaltNode(ctrl, frameptr, reason
2155                             PRODUCT_ONLY(COMMA generate_code_in_product));
2156   halt = _gvn.transform(halt);
2157   root()->add_req(halt);



2158 }
2159 
2160 //------------------------------uncommon_trap----------------------------------
2161 // Bail out to the interpreter in mid-method.  Implemented by calling the
2162 // uncommon_trap blob.  This helper function inserts a runtime call with the
2163 // right debug info.
2164 Node* GraphKit::uncommon_trap(int trap_request,
2165                              ciKlass* klass, const char* comment,
2166                              bool must_throw,
2167                              bool keep_exact_action) {
2168   if (failing_internal()) {
2169     stop();
2170   }
2171   if (stopped())  return nullptr; // trap reachable?
2172 
2173   // Note:  If ProfileTraps is true, and if a deopt. actually
2174   // occurs here, the runtime will make sure an MDO exists.  There is
2175   // no need to call method()->ensure_method_data() at this point.
2176 
2177   // Set the stack pointer to the right value for reexecution:

2319  *
2320  * @param n          node that the type applies to
2321  * @param exact_kls  type from profiling
2322  * @param maybe_null did profiling see null?
2323  *
2324  * @return           node with improved type
2325  */
2326 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2327   const Type* current_type = _gvn.type(n);
2328   assert(UseTypeSpeculation, "type speculation must be on");
2329 
2330   const TypePtr* speculative = current_type->speculative();
2331 
2332   // Should the klass from the profile be recorded in the speculative type?
2333   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2334     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2335     const TypeOopPtr* xtype = tklass->as_instance_type();
2336     assert(xtype->klass_is_exact(), "Should be exact");
2337     // Any reason to believe n is not null (from this profiling or a previous one)?
2338     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2339     const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2340     // record the new speculative type's depth
2341     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2342     speculative = speculative->with_inline_depth(jvms()->depth());
2343   } else if (current_type->would_improve_ptr(ptr_kind)) {
2344     // Profiling report that null was never seen so we can change the
2345     // speculative type to non null ptr.
2346     if (ptr_kind == ProfileAlwaysNull) {
2347       speculative = TypePtr::NULL_PTR;
2348     } else {
2349       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2350       const TypePtr* ptr = TypePtr::NOTNULL;
2351       if (speculative != nullptr) {
2352         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2353       } else {
2354         speculative = ptr;
2355       }
2356     }
2357   }
2358 
2359   if (speculative != current_type->speculative()) {
2360     // Build a type with a speculative type (what we think we know
2361     // about the type but will need a guard when we use it)
2362     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2363     // We're changing the type, we need a new CheckCast node to carry
2364     // the new type. The new type depends on the control: what
2365     // profiling tells us is only valid from here as far as we can
2366     // tell.
2367     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2368     cast = _gvn.transform(cast);
2369     replace_in_map(n, cast);
2370     n = cast;
2371   }
2372 
2373   return n;
2374 }
2375 
2376 /**
2377  * Record profiling data from receiver profiling at an invoke with the
2378  * type system so that it can propagate it (speculation)
2379  *
2380  * @param n  receiver node
2381  *
2382  * @return   node with improved type
2383  */
2384 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2385   if (!UseTypeSpeculation) {
2386     return n;
2387   }
2388   ciKlass* exact_kls = profile_has_unique_klass();
2389   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2390   if ((java_bc() == Bytecodes::_checkcast ||
2391        java_bc() == Bytecodes::_instanceof ||
2392        java_bc() == Bytecodes::_aastore) &&
2393       method()->method_data()->is_mature()) {
2394     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2395     if (data != nullptr) {
2396       if (!data->as_BitData()->null_seen()) {
2397         ptr_kind = ProfileNeverNull;







2398       } else {
2399         if (TypeProfileCasts) {
2400           assert(data->is_ReceiverTypeData(), "bad profile data type");
2401           ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2402           uint i = 0;
2403           for (; i < call->row_limit(); i++) {
2404             ciKlass* receiver = call->receiver(i);
2405             if (receiver != nullptr) {
2406               break;




2407             }

2408           }
2409           ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2410         }
2411       }
2412     }
2413   }
2414   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2415 }
2416 
2417 /**
2418  * Record profiling data from argument profiling at an invoke with the
2419  * type system so that it can propagate it (speculation)
2420  *
2421  * @param dest_method  target method for the call
2422  * @param bc           what invoke bytecode is this?
2423  */
2424 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2425   if (!UseTypeSpeculation) {
2426     return;
2427   }
2428   const TypeFunc* tf    = TypeFunc::make(dest_method);
2429   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2430   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2431   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2432     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2433     if (is_reference_type(targ->basic_type())) {
2434       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2435       ciKlass* better_type = nullptr;
2436       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2437         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2438       }
2439       i++;
2440     }
2441   }
2442 }
2443 
2444 /**
2445  * Record profiling data from parameter profiling at an invoke with
2446  * the type system so that it can propagate it (speculation)
2447  */
2448 void GraphKit::record_profiled_parameters_for_speculation() {
2449   if (!UseTypeSpeculation) {
2450     return;
2451   }
2452   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {

2572                                   // The first null ends the list.
2573                                   Node* parm0, Node* parm1,
2574                                   Node* parm2, Node* parm3,
2575                                   Node* parm4, Node* parm5,
2576                                   Node* parm6, Node* parm7) {
2577   assert(call_addr != nullptr, "must not call null targets");
2578 
2579   // Slow-path call
2580   bool is_leaf = !(flags & RC_NO_LEAF);
2581   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2582   if (call_name == nullptr) {
2583     assert(!is_leaf, "must supply name for leaf");
2584     call_name = OptoRuntime::stub_name(call_addr);
2585   }
2586   CallNode* call;
2587   if (!is_leaf) {
2588     call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2589   } else if (flags & RC_NO_FP) {
2590     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2591   } else  if (flags & RC_VECTOR){
2592     uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2593     call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2594   } else if (flags & RC_PURE) {
2595     assert(adr_type == nullptr, "pure call does not touch memory");
2596     call = new CallLeafPureNode(call_type, call_addr, call_name);
2597   } else {
2598     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2599   }
2600 
2601   // The following is similar to set_edges_for_java_call,
2602   // except that the memory effects of the call are restricted to AliasIdxRaw.
2603 
2604   // Slow path call has no side-effects, uses few values
2605   bool wide_in  = !(flags & RC_NARROW_MEM);
2606   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2607 
2608   Node* prev_mem = nullptr;
2609   if (wide_in) {
2610     prev_mem = set_predefined_input_for_runtime_call(call);
2611   } else {
2612     assert(!wide_out, "narrow in => narrow out");
2613     Node* narrow_mem = memory(adr_type);
2614     prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2615   }
2616 
2617   // Hook each parm in order.  Stop looking at the first null.
2618   if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2619   if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2620   if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2621   if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2622   if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2623   if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2624   if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2625   if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2626   /* close each nested if ===> */  } } } } } } } }
2627   assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2628 
2629   if (!is_leaf) {
2630     // Non-leaves can block and take safepoints:
2631     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2632   }
2633   // Non-leaves can throw exceptions:
2634   if (has_io) {
2635     call->set_req(TypeFunc::I_O, i_o());
2636   }
2637 
2638   if (flags & RC_UNCOMMON) {
2639     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2640     // (An "if" probability corresponds roughly to an unconditional count.
2641     // Sort of.)
2642     call->set_cnt(PROB_UNLIKELY_MAG(4));
2643   }
2644 
2645   Node* c = _gvn.transform(call);
2646   assert(c == call, "cannot disappear");
2647 

2655 
2656   if (has_io) {
2657     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2658   }
2659   return call;
2660 
2661 }
2662 
2663 // i2b
2664 Node* GraphKit::sign_extend_byte(Node* in) {
2665   Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2666   return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2667 }
2668 
2669 // i2s
2670 Node* GraphKit::sign_extend_short(Node* in) {
2671   Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2672   return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2673 }
2674 

2675 //------------------------------merge_memory-----------------------------------
2676 // Merge memory from one path into the current memory state.
2677 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2678   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2679     Node* old_slice = mms.force_memory();
2680     Node* new_slice = mms.memory2();
2681     if (old_slice != new_slice) {
2682       PhiNode* phi;
2683       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2684         if (mms.is_empty()) {
2685           // clone base memory Phi's inputs for this memory slice
2686           assert(old_slice == mms.base_memory(), "sanity");
2687           phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2688           _gvn.set_type(phi, Type::MEMORY);
2689           for (uint i = 1; i < phi->req(); i++) {
2690             phi->init_req(i, old_slice->in(i));
2691           }
2692         } else {
2693           phi = old_slice->as_Phi(); // Phi was generated already
2694         }

2751   gvn.transform(iff);
2752   if (!bol->is_Con()) gvn.record_for_igvn(iff);
2753   return iff;
2754 }
2755 
2756 //-------------------------------gen_subtype_check-----------------------------
2757 // Generate a subtyping check.  Takes as input the subtype and supertype.
2758 // Returns 2 values: sets the default control() to the true path and returns
2759 // the false path.  Only reads invariant memory; sets no (visible) memory.
2760 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2761 // but that's not exposed to the optimizer.  This call also doesn't take in an
2762 // Object; if you wish to check an Object you need to load the Object's class
2763 // prior to coming here.
2764 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2765                                ciMethod* method, int bci) {
2766   Compile* C = gvn.C;
2767   if ((*ctrl)->is_top()) {
2768     return C->top();
2769   }
2770 








2771   // Fast check for identical types, perhaps identical constants.
2772   // The types can even be identical non-constants, in cases
2773   // involving Array.newInstance, Object.clone, etc.
2774   if (subklass == superklass)
2775     return C->top();             // false path is dead; no test needed.
2776 
2777   if (gvn.type(superklass)->singleton()) {
2778     const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2779     const TypeKlassPtr* subk   = gvn.type(subklass)->is_klassptr();
2780 
2781     // In the common case of an exact superklass, try to fold up the
2782     // test before generating code.  You may ask, why not just generate
2783     // the code and then let it fold up?  The answer is that the generated
2784     // code will necessarily include null checks, which do not always
2785     // completely fold away.  If they are also needless, then they turn
2786     // into a performance loss.  Example:
2787     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2788     // Here, the type of 'fa' is often exact, so the store check
2789     // of fa[1]=x will fold up, without testing the nullness of x.
2790     //
2791     // At macro expansion, we would have already folded the SubTypeCheckNode
2792     // being expanded here because we always perform the static sub type
2793     // check in SubTypeCheckNode::sub() regardless of whether
2794     // StressReflectiveCode is set or not. We can therefore skip this
2795     // static check when StressReflectiveCode is on.
2796     switch (C->static_subtype_check(superk, subk)) {
2797     case Compile::SSC_always_false:
2798       {
2799         Node* always_fail = *ctrl;
2800         *ctrl = gvn.C->top();
2801         return always_fail;
2802       }
2803     case Compile::SSC_always_true:
2804       return C->top();
2805     case Compile::SSC_easy_test:
2806       {
2807         // Just do a direct pointer compare and be done.
2808         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2809         *ctrl = gvn.transform(new IfTrueNode(iff));
2810         return gvn.transform(new IfFalseNode(iff));
2811       }
2812     case Compile::SSC_full_test:
2813       break;
2814     default:
2815       ShouldNotReachHere();
2816     }
2817   }
2818 
2819   // %%% Possible further optimization:  Even if the superklass is not exact,
2820   // if the subklass is the unique subtype of the superklass, the check
2821   // will always succeed.  We could leave a dependency behind to ensure this.
2822 
2823   // First load the super-klass's check-offset
2824   Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2825   Node* m = C->immutable_memory();
2826   Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2827   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2828   const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();

2866   gvn.record_for_igvn(r_ok_subtype);
2867 
2868   // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2869   // SubTypeCheck node
2870   if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2871     ciCallProfile profile = method->call_profile_at_bci(bci);
2872     float total_prob = 0;
2873     for (int i = 0; profile.has_receiver(i); ++i) {
2874       float prob = profile.receiver_prob(i);
2875       total_prob += prob;
2876     }
2877     if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2878       const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2879       for (int i = 0; profile.has_receiver(i); ++i) {
2880         ciKlass* klass = profile.receiver(i);
2881         const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2882         Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2883         if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2884           continue;
2885         }




2886         float prob = profile.receiver_prob(i);
2887         ConNode* klass_node = gvn.makecon(klass_t);
2888         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2889         Node* iftrue = gvn.transform(new IfTrueNode(iff));
2890 
2891         if (result == Compile::SSC_always_true) {
2892           r_ok_subtype->add_req(iftrue);
2893         } else {
2894           assert(result == Compile::SSC_always_false, "");
2895           r_not_subtype->add_req(iftrue);
2896         }
2897         *ctrl = gvn.transform(new IfFalseNode(iff));
2898       }
2899     }
2900   }
2901 
2902   // See if we get an immediate positive hit.  Happens roughly 83% of the
2903   // time.  Test to see if the value loaded just previously from the subklass
2904   // is exactly the superklass.
2905   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);

2919       igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
2920     }
2921     return not_subtype_ctrl;
2922   }
2923 
2924   r_ok_subtype->init_req(1, iftrue1);
2925 
2926   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2927   // is roughly 63% of the remaining cases).  Test to see if the loaded
2928   // check-offset points into the subklass display list or the 1-element
2929   // cache.  If it points to the display (and NOT the cache) and the display
2930   // missed then it's not a subtype.
2931   Node *cacheoff = gvn.intcon(cacheoff_con);
2932   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2933   r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2934   *ctrl = gvn.transform(new IfFalseNode(iff2));
2935 
2936   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2937   // No performance impact (too rare) but allows sharing of secondary arrays
2938   // which has some footprint reduction.
2939   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2940   r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2941   *ctrl = gvn.transform(new IfFalseNode(iff3));
2942 
2943   // -- Roads not taken here: --
2944   // We could also have chosen to perform the self-check at the beginning
2945   // of this code sequence, as the assembler does.  This would not pay off
2946   // the same way, since the optimizer, unlike the assembler, can perform
2947   // static type analysis to fold away many successful self-checks.
2948   // Non-foldable self checks work better here in second position, because
2949   // the initial primary superclass check subsumes a self-check for most
2950   // types.  An exception would be a secondary type like array-of-interface,
2951   // which does not appear in its own primary supertype display.
2952   // Finally, we could have chosen to move the self-check into the
2953   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2954   // dependent manner.  But it is worthwhile to have the check here,
2955   // where it can be perhaps be optimized.  The cost in code space is
2956   // small (register compare, branch).
2957 
2958   // Now do a linear scan of the secondary super-klass array.  Again, no real
2959   // performance impact (too rare) but it's gotta be done.
2960   // Since the code is rarely used, there is no penalty for moving it
2961   // out of line, and it can only improve I-cache density.
2962   // The decision to inline or out-of-line this final check is platform
2963   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2964   Node* psc = gvn.transform(
2965     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2966 
2967   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2968   r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2969   r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2970 
2971   // Return false path; set default control to true path.
2972   *ctrl = gvn.transform(r_ok_subtype);
2973   return gvn.transform(r_not_subtype);
2974 }
2975 
2976 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {





2977   bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2978   if (expand_subtype_check) {
2979     MergeMemNode* mem = merged_memory();
2980     Node* ctrl = control();
2981     Node* subklass = obj_or_subklass;
2982     if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2983       subklass = load_object_klass(obj_or_subklass);
2984     }
2985 
2986     Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2987     set_control(ctrl);
2988     return n;
2989   }
2990 
2991   Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2992   Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2993   IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2994   set_control(_gvn.transform(new IfTrueNode(iff)));
2995   return _gvn.transform(new IfFalseNode(iff));
2996 }
2997 
2998 // Profile-driven exact type check:
2999 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3000                                     float prob,
3001                                     Node* *casted_receiver) {
3002   assert(!klass->is_interface(), "no exact type check on interfaces");
3003 











3004   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);




3005   Node* recv_klass = load_object_klass(receiver);
3006   Node* want_klass = makecon(tklass);
3007   Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3008   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3009   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3010   set_control( _gvn.transform(new IfTrueNode (iff)));
3011   Node* fail = _gvn.transform(new IfFalseNode(iff));
3012 
3013   if (!stopped()) {
3014     const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3015     const TypeOopPtr* recvx_type = tklass->as_instance_type();
3016     assert(recvx_type->klass_is_exact(), "");
3017 
3018     if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
3019       // Subsume downstream occurrences of receiver with a cast to
3020       // recv_xtype, since now we know what the type will be.
3021       Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
3022       (*casted_receiver) = _gvn.transform(cast);





3023       assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3024       // (User must make the replace_in_map call.)
3025     }
3026   }
3027 
3028   return fail;
3029 }
3030 











3031 //------------------------------subtype_check_receiver-------------------------
3032 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3033                                        Node** casted_receiver) {
3034   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3035   Node* want_klass = makecon(tklass);
3036 
3037   Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3038 
3039   // Ignore interface type information until interface types are properly tracked.
3040   if (!stopped() && !klass->is_interface()) {
3041     const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3042     const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3043     if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3044       Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3045       (*casted_receiver) = _gvn.transform(cast);



3046     }
3047   }
3048 
3049   return slow_ctl;
3050 }
3051 
3052 //------------------------------seems_never_null-------------------------------
3053 // Use null_seen information if it is available from the profile.
3054 // If we see an unexpected null at a type check we record it and force a
3055 // recompile; the offending check will be recompiled to handle nulls.
3056 // If we see several offending BCIs, then all checks in the
3057 // method will be recompiled.
3058 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3059   speculating = !_gvn.type(obj)->speculative_maybe_null();
3060   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3061   if (UncommonNullCast               // Cutout for this technique
3062       && obj != null()               // And not the -Xcomp stupid case?
3063       && !too_many_traps(reason)
3064       ) {
3065     if (speculating) {

3134 
3135 //------------------------maybe_cast_profiled_receiver-------------------------
3136 // If the profile has seen exactly one type, narrow to exactly that type.
3137 // Subsequent type checks will always fold up.
3138 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3139                                              const TypeKlassPtr* require_klass,
3140                                              ciKlass* spec_klass,
3141                                              bool safe_for_replace) {
3142   if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3143 
3144   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3145 
3146   // Make sure we haven't already deoptimized from this tactic.
3147   if (too_many_traps_or_recompiles(reason))
3148     return nullptr;
3149 
3150   // (No, this isn't a call, but it's enough like a virtual call
3151   // to use the same ciMethod accessor to get the profile info...)
3152   // If we have a speculative type use it instead of profiling (which
3153   // may not help us)
3154   ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;













3155   if (exact_kls != nullptr) {// no cast failures here
3156     if (require_klass == nullptr ||
3157         C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3158       // If we narrow the type to match what the type profile sees or
3159       // the speculative type, we can then remove the rest of the
3160       // cast.
3161       // This is a win, even if the exact_kls is very specific,
3162       // because downstream operations, such as method calls,
3163       // will often benefit from the sharper type.
3164       Node* exact_obj = not_null_obj; // will get updated in place...
3165       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
3166                                             &exact_obj);
3167       { PreserveJVMState pjvms(this);
3168         set_control(slow_ctl);
3169         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3170       }
3171       if (safe_for_replace) {
3172         replace_in_map(not_null_obj, exact_obj);
3173       }
3174       return exact_obj;

3264   // If not_null_obj is dead, only null-path is taken
3265   if (stopped()) {              // Doing instance-of on a null?
3266     set_control(null_ctl);
3267     return intcon(0);
3268   }
3269   region->init_req(_null_path, null_ctl);
3270   phi   ->init_req(_null_path, intcon(0)); // Set null path value
3271   if (null_ctl == top()) {
3272     // Do this eagerly, so that pattern matches like is_diamond_phi
3273     // will work even during parsing.
3274     assert(_null_path == PATH_LIMIT-1, "delete last");
3275     region->del_req(_null_path);
3276     phi   ->del_req(_null_path);
3277   }
3278 
3279   // Do we know the type check always succeed?
3280   bool known_statically = false;
3281   if (_gvn.type(superklass)->singleton()) {
3282     const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3283     const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3284     if (subk->is_loaded()) {
3285       int static_res = C->static_subtype_check(superk, subk);
3286       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3287     }
3288   }
3289 
3290   if (!known_statically) {
3291     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3292     // We may not have profiling here or it may not help us. If we
3293     // have a speculative type use it to perform an exact cast.
3294     ciKlass* spec_obj_type = obj_type->speculative_type();
3295     if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3296       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3297       if (stopped()) {            // Profile disagrees with this path.
3298         set_control(null_ctl);    // Null is the only remaining possibility.
3299         return intcon(0);
3300       }
3301       if (cast_obj != nullptr) {
3302         not_null_obj = cast_obj;
3303       }
3304     }

3320   record_for_igvn(region);
3321 
3322   // If we know the type check always succeeds then we don't use the
3323   // profiling data at this bytecode. Don't lose it, feed it to the
3324   // type system as a speculative type.
3325   if (safe_for_replace) {
3326     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3327     replace_in_map(obj, casted_obj);
3328   }
3329 
3330   return _gvn.transform(phi);
3331 }
3332 
3333 //-------------------------------gen_checkcast---------------------------------
3334 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3335 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3336 // uncommon-trap paths work.  Adjust stack after this call.
3337 // If failure_control is supplied and not null, it is filled in with
3338 // the control edge for the cast failure.  Otherwise, an appropriate
3339 // uncommon trap or exception is thrown.
3340 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3341                               Node* *failure_control) {





3342   kill_dead_locals();           // Benefit all the uncommon traps
3343   const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();


3344   const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3345   const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();


3346 
3347   // Fast cutout:  Check the case that the cast is vacuously true.
3348   // This detects the common cases where the test will short-circuit
3349   // away completely.  We do this before we perform the null check,
3350   // because if the test is going to turn into zero code, we don't
3351   // want a residual null check left around.  (Causes a slowdown,
3352   // for example, in some objArray manipulations, such as a[i]=a[j].)
3353   if (improved_klass_ptr_type->singleton()) {
3354     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3355     if (objtp != nullptr) {
3356       switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {







3357       case Compile::SSC_always_true:
3358         // If we know the type check always succeed then we don't use
3359         // the profiling data at this bytecode. Don't lose it, feed it
3360         // to the type system as a speculative type.
3361         return record_profiled_receiver_for_speculation(obj);





3362       case Compile::SSC_always_false:




3363         // It needs a null check because a null will *pass* the cast check.
3364         // A non-null value will always produce an exception.
3365         if (!objtp->maybe_null()) {
3366           bool is_aastore = (java_bc() == Bytecodes::_aastore);
3367           Deoptimization::DeoptReason reason = is_aastore ?
3368             Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3369           builtin_throw(reason);
3370           return top();
3371         } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3372           return null_assert(obj);
3373         }
3374         break; // Fall through to full check
3375       default:
3376         break;
3377       }
3378     }
3379   }
3380 
3381   ciProfileData* data = nullptr;
3382   bool safe_for_replace = false;
3383   if (failure_control == nullptr) {        // use MDO in regular case only
3384     assert(java_bc() == Bytecodes::_aastore ||
3385            java_bc() == Bytecodes::_checkcast,
3386            "interpreter profiles type checks only for these BCs");
3387     data = method()->method_data()->bci_to_data(bci());
3388     safe_for_replace = true;

3389   }
3390 
3391   // Make the merge point
3392   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3393   RegionNode* region = new RegionNode(PATH_LIMIT);
3394   Node*       phi    = new PhiNode(region, toop);



3395   C->set_has_split_ifs(true); // Has chance for split-if optimization
3396 
3397   // Use null-cast information if it is available
3398   bool speculative_not_null = false;
3399   bool never_see_null = ((failure_control == nullptr)  // regular case only
3400                          && seems_never_null(obj, data, speculative_not_null));
3401 
3402   // Null check; get casted pointer; set region slot 3
3403   Node* null_ctl = top();
3404   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, false /*safe_for_replace*/, speculative_not_null);






3405 
3406   // If not_null_obj is dead, only null-path is taken
3407   if (stopped()) {              // Doing instance-of on a null?
3408     set_control(null_ctl);



3409     return null();
3410   }
3411   region->init_req(_null_path, null_ctl);
3412   phi   ->init_req(_null_path, null());  // Set null path value
3413   if (null_ctl == top()) {
3414     // Do this eagerly, so that pattern matches like is_diamond_phi
3415     // will work even during parsing.
3416     assert(_null_path == PATH_LIMIT-1, "delete last");
3417     region->del_req(_null_path);
3418     phi   ->del_req(_null_path);
3419   }
3420 
3421   Node* cast_obj = nullptr;
3422   if (improved_klass_ptr_type->klass_is_exact()) {
3423     // The following optimization tries to statically cast the speculative type of the object
3424     // (for example obtained during profiling) to the type of the superklass and then do a
3425     // dynamic check that the type of the object is what we expect. To work correctly
3426     // for checkcast and aastore the type of superklass should be exact.
3427     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3428     // We may not have profiling here or it may not help us. If we have
3429     // a speculative type use it to perform an exact cast.
3430     ciKlass* spec_obj_type = obj_type->speculative_type();
3431     if (spec_obj_type != nullptr || data != nullptr) {
3432       cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, false /*safe_for_replace*/);
3433       if (cast_obj != nullptr) {
3434         if (failure_control != nullptr) // failure is now impossible
3435           (*failure_control) = top();
3436         // adjust the type of the phi to the exact klass:
3437         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3438       }
3439     }
3440   }
3441 
3442   if (cast_obj == nullptr) {
3443     // Generate the subtype check
3444     Node* improved_superklass = superklass;
3445     if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {



3446       improved_superklass = makecon(improved_klass_ptr_type);
3447     }
3448     Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3449 
3450     // Plug in success path into the merge
3451     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3452     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3453     if (failure_control == nullptr) {
3454       if (not_subtype_ctrl != top()) { // If failure is possible
3455         PreserveJVMState pjvms(this);
3456         set_control(not_subtype_ctrl);
3457         bool is_aastore = (java_bc() == Bytecodes::_aastore);
3458         Deoptimization::DeoptReason reason = is_aastore ?
3459           Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3460         builtin_throw(reason);
3461       }
3462     } else {
3463       (*failure_control) = not_subtype_ctrl;
3464     }
3465   }
3466 
3467   region->init_req(_obj_path, control());
3468   phi   ->init_req(_obj_path, cast_obj);
3469 
3470   // Return final merged results
3471   set_control( _gvn.transform(region) );
3472   record_for_igvn(region);
3473 


















































3474   // A merge of null or Casted-NotNull obj
3475   Node* res = _gvn.transform(phi);
3476   res = record_profiled_receiver_for_speculation(res);
3477   if (safe_for_replace) {
3478     replace_in_map(obj, res);










3479   }
3480   return res;
3481 }
3482 





































































































































3483 //------------------------------next_monitor-----------------------------------
3484 // What number should be given to the next monitor?
3485 int GraphKit::next_monitor() {
3486   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3487   int next = current + C->sync_stack_slots();
3488   // Keep the toplevel high water mark current:
3489   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3490   return current;
3491 }
3492 
3493 //------------------------------insert_mem_bar---------------------------------
3494 // Memory barrier to avoid floating things around
3495 // The membar serves as a pinch point between both control and all memory slices.
3496 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3497   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3498   mb->init_req(TypeFunc::Control, control());
3499   mb->init_req(TypeFunc::Memory,  reset_memory());
3500   Node* membar = _gvn.transform(mb);
3501   record_for_igvn(membar);
3502   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));

3603     lock->create_lock_counter(map()->jvms());
3604     increment_counter(lock->counter()->addr());
3605   }
3606 #endif
3607 
3608   return flock;
3609 }
3610 
3611 
3612 //------------------------------shared_unlock----------------------------------
3613 // Emit unlocking code.
3614 void GraphKit::shared_unlock(Node* box, Node* obj) {
3615   // bci is either a monitorenter bc or InvocationEntryBci
3616   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3617   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3618 
3619   if (stopped()) {               // Dead monitor?
3620     map()->pop_monitor();        // Kill monitor from debug info
3621     return;
3622   }

3623 
3624   // Memory barrier to avoid floating things down past the locked region
3625   insert_mem_bar(Op_MemBarReleaseLock);
3626 
3627   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3628   UnlockNode *unlock = new UnlockNode(C, tf);
3629 #ifdef ASSERT
3630   unlock->set_dbg_jvms(sync_jvms());
3631 #endif
3632   uint raw_idx = Compile::AliasIdxRaw;
3633   unlock->init_req( TypeFunc::Control, control() );
3634   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3635   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3636   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3637   unlock->init_req( TypeFunc::ReturnAdr, top() );
3638 
3639   unlock->init_req(TypeFunc::Parms + 0, obj);
3640   unlock->init_req(TypeFunc::Parms + 1, box);
3641   unlock = _gvn.transform(unlock)->as_Unlock();
3642 
3643   Node* mem = reset_memory();
3644 
3645   // unlock has no side-effects, sets few values
3646   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3647 
3648   // Kill monitor from debug info
3649   map()->pop_monitor( );
3650 }
3651 
3652 //-------------------------------get_layout_helper-----------------------------
3653 // If the given klass is a constant or known to be an array,
3654 // fetch the constant layout helper value into constant_value
3655 // and return null.  Otherwise, load the non-constant
3656 // layout helper value, and return the node which represents it.
3657 // This two-faced routine is useful because allocation sites
3658 // almost always feature constant types.
3659 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3660   const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3661   if (!StressReflectiveCode && klass_t != nullptr) {
3662     bool xklass = klass_t->klass_is_exact();
3663     if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {







3664       jint lhelper;
3665       if (klass_t->isa_aryklassptr()) {
3666         BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();


3667         if (is_reference_type(elem, true)) {
3668           elem = T_OBJECT;
3669         }
3670         lhelper = Klass::array_layout_helper(elem);
3671       } else {
3672         lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3673       }
3674       if (lhelper != Klass::_lh_neutral_value) {
3675         constant_value = lhelper;
3676         return (Node*) nullptr;
3677       }
3678     }
3679   }
3680   constant_value = Klass::_lh_neutral_value;  // put in a known value
3681   Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
3682   return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3683 }
3684 
3685 // We just put in an allocate/initialize with a big raw-memory effect.
3686 // Hook selected additional alias categories on the initialization.
3687 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3688                                 MergeMemNode* init_in_merge,
3689                                 Node* init_out_raw) {
3690   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3691   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3692 
3693   Node* prevmem = kit.memory(alias_idx);
3694   init_in_merge->set_memory_at(alias_idx, prevmem);
3695   kit.set_memory(init_out_raw, alias_idx);


3696 }
3697 
3698 //---------------------------set_output_for_allocation-------------------------
3699 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3700                                           const TypeOopPtr* oop_type,
3701                                           bool deoptimize_on_exception) {
3702   int rawidx = Compile::AliasIdxRaw;
3703   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3704   add_safepoint_edges(alloc);
3705   Node* allocx = _gvn.transform(alloc);
3706   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3707   // create memory projection for i_o
3708   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3709   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3710 
3711   // create a memory projection as for the normal control path
3712   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3713   set_memory(malloc, rawidx);
3714 
3715   // a normal slow-call doesn't change i_o, but an allocation does
3716   // we create a separate i_o projection for the normal control path
3717   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3718   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3719 
3720   // put in an initialization barrier
3721   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3722                                                  rawoop)->as_Initialize();
3723   assert(alloc->initialization() == init,  "2-way macro link must work");
3724   assert(init ->allocation()     == alloc, "2-way macro link must work");
3725   {
3726     // Extract memory strands which may participate in the new object's
3727     // initialization, and source them from the new InitializeNode.
3728     // This will allow us to observe initializations when they occur,
3729     // and link them properly (as a group) to the InitializeNode.
3730     assert(init->in(InitializeNode::Memory) == malloc, "");
3731     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3732     init->set_req(InitializeNode::Memory, minit_in);
3733     record_for_igvn(minit_in); // fold it up later, if possible

3734     Node* minit_out = memory(rawidx);
3735     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3736     int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
3737     // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
3738     // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
3739     // multiple projections as a result.
3740     set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
3741     int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
3742     set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
3743     if (oop_type->isa_aryptr()) {





3744       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3745       int            elemidx  = C->get_alias_index(telemref);
3746       hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(elemidx))));




3747     } else if (oop_type->isa_instptr()) {
3748       ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3749       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3750         ciField* field = ik->nonstatic_field_at(i);
3751         if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3752           continue;  // do not bother to track really large numbers of fields
3753         // Find (or create) the alias category for this field:
3754         int fieldidx = C->alias_type(field)->index();
3755         hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
3756       }
3757     }
3758   }
3759 
3760   // Cast raw oop to the real thing...
3761   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3762   javaoop = _gvn.transform(javaoop);
3763   C->set_recent_alloc(control(), javaoop);
3764   assert(just_allocated_object(control()) == javaoop, "just allocated");
3765 
3766 #ifdef ASSERT

3778       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3779     }
3780   }
3781 #endif //ASSERT
3782 
3783   return javaoop;
3784 }
3785 
3786 //---------------------------new_instance--------------------------------------
3787 // This routine takes a klass_node which may be constant (for a static type)
3788 // or may be non-constant (for reflective code).  It will work equally well
3789 // for either, and the graph will fold nicely if the optimizer later reduces
3790 // the type to a constant.
3791 // The optional arguments are for specialized use by intrinsics:
3792 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3793 //  - If 'return_size_val', report the total object size to the caller.
3794 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3795 Node* GraphKit::new_instance(Node* klass_node,
3796                              Node* extra_slow_test,
3797                              Node* *return_size_val,
3798                              bool deoptimize_on_exception) {

3799   // Compute size in doublewords
3800   // The size is always an integral number of doublewords, represented
3801   // as a positive bytewise size stored in the klass's layout_helper.
3802   // The layout_helper also encodes (in a low bit) the need for a slow path.
3803   jint  layout_con = Klass::_lh_neutral_value;
3804   Node* layout_val = get_layout_helper(klass_node, layout_con);
3805   int   layout_is_con = (layout_val == nullptr);
3806 
3807   if (extra_slow_test == nullptr)  extra_slow_test = intcon(0);
3808   // Generate the initial go-slow test.  It's either ALWAYS (return a
3809   // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3810   // case) a computed value derived from the layout_helper.
3811   Node* initial_slow_test = nullptr;
3812   if (layout_is_con) {
3813     assert(!StressReflectiveCode, "stress mode does not use these paths");
3814     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3815     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3816   } else {   // reflective case
3817     // This reflective path is used by Unsafe.allocateInstance.
3818     // (It may be stress-tested by specifying StressReflectiveCode.)
3819     // Basically, we want to get into the VM is there's an illegal argument.
3820     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3821     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3822     if (extra_slow_test != intcon(0)) {
3823       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3824     }
3825     // (Macro-expander will further convert this to a Bool, if necessary.)

3836 
3837     // Clear the low bits to extract layout_helper_size_in_bytes:
3838     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3839     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3840     size = _gvn.transform( new AndXNode(size, mask) );
3841   }
3842   if (return_size_val != nullptr) {
3843     (*return_size_val) = size;
3844   }
3845 
3846   // This is a precise notnull oop of the klass.
3847   // (Actually, it need not be precise if this is a reflective allocation.)
3848   // It's what we cast the result to.
3849   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3850   if (!tklass)  tklass = TypeInstKlassPtr::OBJECT;
3851   const TypeOopPtr* oop_type = tklass->as_instance_type();
3852 
3853   // Now generate allocation code
3854 
3855   // The entire memory state is needed for slow path of the allocation
3856   // since GC and deoptimization can happened.
3857   Node *mem = reset_memory();
3858   set_all_memory(mem); // Create new memory state
3859 
3860   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3861                                          control(), mem, i_o(),
3862                                          size, klass_node,
3863                                          initial_slow_test);
3864 
3865   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3866 }
3867 
3868 //-------------------------------new_array-------------------------------------
3869 // helper for both newarray and anewarray
3870 // The 'length' parameter is (obviously) the length of the array.
3871 // The optional arguments are for specialized use by intrinsics:
3872 //  - If 'return_size_val', report the non-padded array size (sum of header size
3873 //    and array body) to the caller.
3874 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3875 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3876                           Node* length,         // number of array elements
3877                           int   nargs,          // number of arguments to push back for uncommon trap
3878                           Node* *return_size_val,
3879                           bool deoptimize_on_exception) {

3880   jint  layout_con = Klass::_lh_neutral_value;
3881   Node* layout_val = get_layout_helper(klass_node, layout_con);
3882   int   layout_is_con = (layout_val == nullptr);
3883 
3884   if (!layout_is_con && !StressReflectiveCode &&
3885       !too_many_traps(Deoptimization::Reason_class_check)) {
3886     // This is a reflective array creation site.
3887     // Optimistically assume that it is a subtype of Object[],
3888     // so that we can fold up all the address arithmetic.
3889     layout_con = Klass::array_layout_helper(T_OBJECT);
3890     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3891     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3892     { BuildCutout unless(this, bol_lh, PROB_MAX);
3893       inc_sp(nargs);
3894       uncommon_trap(Deoptimization::Reason_class_check,
3895                     Deoptimization::Action_maybe_recompile);
3896     }
3897     layout_val = nullptr;
3898     layout_is_con = true;
3899   }
3900 
3901   // Generate the initial go-slow test.  Make sure we do not overflow
3902   // if length is huge (near 2Gig) or negative!  We do not need
3903   // exact double-words here, just a close approximation of needed
3904   // double-words.  We can't add any offset or rounding bits, lest we
3905   // take a size -1 of bytes and make it positive.  Use an unsigned
3906   // compare, so negative sizes look hugely positive.
3907   int fast_size_limit = FastAllocateSizeLimit;
3908   if (layout_is_con) {
3909     assert(!StressReflectiveCode, "stress mode does not use these paths");
3910     // Increase the size limit if we have exact knowledge of array type.
3911     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3912     assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3913            "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3914     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3915   }
3916 
3917   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3918   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3919 
3920   // --- Size Computation ---
3921   // array_size = round_to_heap(array_header + (length << elem_shift));
3922   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3923   // and align_to(x, y) == ((x + y-1) & ~(y-1))
3924   // The rounding mask is strength-reduced, if possible.
3925   int round_mask = MinObjAlignmentInBytes - 1;
3926   Node* header_size = nullptr;
3927   // (T_BYTE has the weakest alignment and size restrictions...)
3928   if (layout_is_con) {
3929     int       hsize  = Klass::layout_helper_header_size(layout_con);
3930     int       eshift = Klass::layout_helper_log2_element_size(layout_con);

3931     if ((round_mask & ~right_n_bits(eshift)) == 0)
3932       round_mask = 0;  // strength-reduce it if it goes away completely
3933     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3934     int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3935     assert(header_size_min <= hsize, "generic minimum is smallest");
3936     header_size = intcon(hsize);
3937   } else {
3938     Node* hss   = intcon(Klass::_lh_header_size_shift);
3939     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3940     header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3941     header_size = _gvn.transform(new AndINode(header_size, hsm));
3942   }
3943 
3944   Node* elem_shift = nullptr;
3945   if (layout_is_con) {
3946     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3947     if (eshift != 0)
3948       elem_shift = intcon(eshift);
3949   } else {
3950     // There is no need to mask or shift this value.
3951     // The semantics of LShiftINode include an implicit mask to 0x1F.
3952     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3953     elem_shift = layout_val;

4002   }
4003   Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4004 
4005   if (return_size_val != nullptr) {
4006     // This is the size
4007     (*return_size_val) = non_rounded_size;
4008   }
4009 
4010   Node* size = non_rounded_size;
4011   if (round_mask != 0) {
4012     Node* mask1 = MakeConX(round_mask);
4013     size = _gvn.transform(new AddXNode(size, mask1));
4014     Node* mask2 = MakeConX(~round_mask);
4015     size = _gvn.transform(new AndXNode(size, mask2));
4016   }
4017   // else if round_mask == 0, the size computation is self-rounding
4018 
4019   // Now generate allocation code
4020 
4021   // The entire memory state is needed for slow path of the allocation
4022   // since GC and deoptimization can happened.
4023   Node *mem = reset_memory();
4024   set_all_memory(mem); // Create new memory state
4025 
4026   if (initial_slow_test->is_Bool()) {
4027     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4028     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4029   }
4030 
4031   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();




















4032   Node* valid_length_test = _gvn.intcon(1);
4033   if (ary_type->isa_aryptr()) {
4034     BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4035     jint max = TypeAryPtr::max_array_length(bt);
4036     Node* valid_length_cmp  = _gvn.transform(new CmpUNode(length, intcon(max)));
4037     valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4038   }
4039 
4040   // Create the AllocateArrayNode and its result projections
4041   AllocateArrayNode* alloc
4042     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4043                             control(), mem, i_o(),
4044                             size, klass_node,
4045                             initial_slow_test,
4046                             length, valid_length_test);
4047 
4048   // Cast to correct type.  Note that the klass_node may be constant or not,
4049   // and in the latter case the actual array type will be inexact also.
4050   // (This happens via a non-constant argument to inline_native_newArray.)
4051   // In any case, the value of klass_node provides the desired array type.
4052   const TypeInt* length_type = _gvn.find_int_type(length);
4053   if (ary_type->isa_aryptr() && length_type != nullptr) {
4054     // Try to get a better type than POS for the size
4055     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4056   }
4057 
4058   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4059 
4060   array_ideal_length(alloc, ary_type, true);
4061   return javaoop;
4062 }
4063 
4064 // The following "Ideal_foo" functions are placed here because they recognize
4065 // the graph shapes created by the functions immediately above.
4066 
4067 //---------------------------Ideal_allocation----------------------------------

4162 void GraphKit::add_parse_predicates(int nargs) {
4163   if (ShortRunningLongLoop) {
4164     // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4165     // walking up from the loop.
4166     add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4167   }
4168   if (UseLoopPredicate) {
4169     add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4170     if (UseProfiledLoopPredicate) {
4171       add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4172     }
4173   }
4174   if (UseAutoVectorizationPredicate) {
4175     add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4176   }
4177   // Loop Limit Check Predicate should be near the loop.
4178   add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4179 }
4180 
4181 void GraphKit::sync_kit(IdealKit& ideal) {

4182   set_all_memory(ideal.merged_memory());
4183   set_i_o(ideal.i_o());
4184   set_control(ideal.ctrl());
4185 }
4186 
4187 void GraphKit::final_sync(IdealKit& ideal) {
4188   // Final sync IdealKit and graphKit.
4189   sync_kit(ideal);
4190 }
4191 
4192 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4193   Node* len = load_array_length(load_String_value(str, set_ctrl));
4194   Node* coder = load_String_coder(str, set_ctrl);
4195   // Divide length by 2 if coder is UTF16
4196   return _gvn.transform(new RShiftINode(len, coder));
4197 }
4198 
4199 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4200   int value_offset = java_lang_String::value_offset();
4201   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4202                                                      false, nullptr, 0);
4203   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4204   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::BotPTR,
4205                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4206                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
4207   Node* p = basic_plus_adr(str, str, value_offset);
4208   Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4209                               IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4210   return must_be_not_null(load, true);
4211 }
4212 
4213 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4214   if (!CompactStrings) {
4215     return intcon(java_lang_String::CODER_UTF16);
4216   }
4217   int coder_offset = java_lang_String::coder_offset();
4218   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4219                                                      false, nullptr, 0);
4220   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4221 
4222   Node* p = basic_plus_adr(str, str, coder_offset);
4223   Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4224                               IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4225   return load;
4226 }
4227 
4228 void GraphKit::store_String_value(Node* str, Node* value) {
4229   int value_offset = java_lang_String::value_offset();
4230   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4231                                                      false, nullptr, 0);
4232   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4233 
4234   access_store_at(str,  basic_plus_adr(str, value_offset), value_field_type,
4235                   value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4236 }
4237 
4238 void GraphKit::store_String_coder(Node* str, Node* value) {
4239   int coder_offset = java_lang_String::coder_offset();
4240   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4241                                                      false, nullptr, 0);
4242   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4243 
4244   access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4245                   value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4246 }
4247 
4248 // If input and output memory types differ, capture the whole memory to preserve
4249 // the dependency between preceding and subsequent loads/stores.
4250 // For example, the following program:
4251 //  StoreB
4252 //  compress_string
4253 //  LoadB
4254 // has this memory graph (use->def):
4255 //  LoadB -> compress_string -> CharMem
4256 //             ... -> StoreB -> ByteMem
4257 // The intrinsic hides the dependency between LoadB and StoreB, causing
4258 // the load to read from memory not containing the result of the StoreB.
4259 // The correct memory graph should look like this:
4260 //  LoadB -> compress_string -> MergeMem -> StoreB
4261 Node* GraphKit::capture_memory(const TypePtr*& combined_type, const TypePtr* src_type, const TypePtr* dst_type) {

4364   i_char->init_req(2, AddI(i_char, intcon(2)));
4365 
4366   set_control(IfFalse(iff));
4367   set_memory(st, TypeAryPtr::BYTES);
4368 }
4369 
4370 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4371   if (!field->is_constant()) {
4372     return nullptr; // Field not marked as constant.
4373   }
4374   ciInstance* holder = nullptr;
4375   if (!field->is_static()) {
4376     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4377     if (const_oop != nullptr && const_oop->is_instance()) {
4378       holder = const_oop->as_instance();
4379     }
4380   }
4381   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4382                                                         /*is_unsigned_load=*/false);
4383   if (con_type != nullptr) {
4384     return makecon(con_type);






4385   }
4386   return nullptr;
4387 }
4388 
4389 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4390   const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4391   const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4392   if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4393     const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4394     Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4395     return casted_obj;



4396   }
4397   return obj;
4398 }

   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "asm/register.hpp"
  26 #include "ci/ciFlatArrayKlass.hpp"
  27 #include "ci/ciInlineKlass.hpp"
  28 #include "ci/ciMethod.hpp"
  29 #include "ci/ciObjArray.hpp"
  30 #include "ci/ciUtilities.hpp"
  31 #include "classfile/javaClasses.hpp"
  32 #include "compiler/compileLog.hpp"
  33 #include "gc/shared/barrierSet.hpp"
  34 #include "gc/shared/c2/barrierSetC2.hpp"
  35 #include "interpreter/interpreter.hpp"
  36 #include "memory/resourceArea.hpp"
  37 #include "oops/flatArrayKlass.hpp"
  38 #include "opto/addnode.hpp"
  39 #include "opto/callnode.hpp"
  40 #include "opto/castnode.hpp"
  41 #include "opto/convertnode.hpp"
  42 #include "opto/graphKit.hpp"
  43 #include "opto/idealKit.hpp"
  44 #include "opto/inlinetypenode.hpp"
  45 #include "opto/intrinsicnode.hpp"
  46 #include "opto/locknode.hpp"
  47 #include "opto/machnode.hpp"
  48 #include "opto/memnode.hpp"
  49 #include "opto/multnode.hpp"
  50 #include "opto/narrowptrnode.hpp"
  51 #include "opto/opaquenode.hpp"
  52 #include "opto/opcodes.hpp"
  53 #include "opto/parse.hpp"
  54 #include "opto/reachability.hpp"
  55 #include "opto/rootnode.hpp"
  56 #include "opto/runtime.hpp"
  57 #include "opto/subtypenode.hpp"
  58 #include "opto/type.hpp"
  59 #include "runtime/arguments.hpp"
  60 #include "runtime/deoptimization.hpp"
  61 #include "runtime/sharedRuntime.hpp"
  62 #include "runtime/stubRoutines.hpp"
  63 #include "utilities/bitMap.inline.hpp"
  64 #include "utilities/growableArray.hpp"
  65 #include "utilities/powerOfTwo.hpp"
  66 
  67 //----------------------------GraphKit-----------------------------------------
  68 // Main utility constructor.
  69 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
  70   : Phase(Phase::Parser),
  71     _env(C->env()),
  72     _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
  73     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  74 {
  75   assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
  76   _exceptions = jvms->map()->next_exception();
  77   if (_exceptions != nullptr)  jvms->map()->set_next_exception(nullptr);
  78   set_jvms(jvms);
  79 #ifdef ASSERT
  80   if (_gvn.is_IterGVN() != nullptr) {
  81     assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
  82     // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
  83     _worklist_size = _gvn.C->igvn_worklist()->size();
  84   }
  85 #endif
  86 }
  87 
  88 // Private constructor for parser.
  89 GraphKit::GraphKit()
  90   : Phase(Phase::Parser),
  91     _env(C->env()),
  92     _gvn(*C->initial_gvn()),
  93     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  94 {
  95   _exceptions = nullptr;
  96   set_map(nullptr);
  97   DEBUG_ONLY(_sp = -99);
  98   DEBUG_ONLY(set_bci(-99));
  99 }
 100 
 101 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
 102   : Phase(Phase::Parser),
 103     _env(C->env()),
 104     _gvn(igvn),
 105     _exceptions(nullptr),
 106     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
 107   assert(igvn.delay_transform(), "must delay transformation during macro expansion");
 108   assert(sft->next_exception() == nullptr, "must not have a pending exception");
 109   JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
 110   SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
 111   for (uint i = 0; i < sft->req(); i++) {
 112     cloned_map->init_req(i, sft->in(i));
 113   }
 114   igvn.record_for_igvn(cloned_map);
 115   for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
 116     current->set_map(cloned_map);
 117   }
 118   set_jvms(cloned_jvms);
 119   set_all_memory(cloned_map->memory());
 120 }
 121 
 122 //---------------------------clean_stack---------------------------------------
 123 // Clear away rubbish from the stack area of the JVM state.
 124 // This destroys any arguments that may be waiting on the stack.
 125 void GraphKit::clean_stack(int from_sp) {
 126   SafePointNode* map      = this->map();
 127   JVMState*      jvms     = this->jvms();
 128   int            stk_size = jvms->stk_size();
 129   int            stkoff   = jvms->stkoff();
 130   Node*          top      = this->top();
 131   for (int i = from_sp; i < stk_size; i++) {
 132     if (map->in(stkoff + i) != top) {
 133       map->set_req(stkoff + i, top);
 134     }
 135   }
 136 }
 137 
 138 
 139 //--------------------------------sync_jvms-----------------------------------
 140 // Make sure our current jvms agrees with our parse state.

 939         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 940           tty->print_cr("Zombie local %d: ", local);
 941           jvms->dump();
 942         }
 943         return false;
 944       }
 945     }
 946   }
 947   return true;
 948 }
 949 
 950 #endif //ASSERT
 951 
 952 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
 953 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 954   ciMethod* cur_method = jvms->method();
 955   int       cur_bci   = jvms->bci();
 956   if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
 957     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 958     return Interpreter::bytecode_should_reexecute(code) ||
 959            (is_anewarray && (code == Bytecodes::_multianewarray));
 960     // Reexecute _multianewarray bytecode which was replaced with
 961     // sequence of [a]newarray. See Parse::do_multianewarray().
 962     //
 963     // Note: interpreter should not have it set since this optimization
 964     // is limited by dimensions and guarded by flag so in some cases
 965     // multianewarray() runtime calls will be generated and
 966     // the bytecode should not be reexecutes (stack will not be reset).
 967   } else {
 968     return false;
 969   }
 970 }
 971 
 972 // Helper function for adding JVMState and debug information to node
 973 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 974   // Add the safepoint edges to the call (or other safepoint).
 975 
 976   // Make sure dead locals are set to top.  This
 977   // should help register allocation time and cut down on the size
 978   // of the deoptimization information.
 979   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");

1055     uint p = debug_start;  // walks forward in [debug_start, debug_end)
1056     uint j, k, l;
1057     SafePointNode* in_map = in_jvms->map();
1058     out_jvms->set_map(call);
1059 
1060     if (can_prune_locals) {
1061       assert(in_jvms->method() == out_jvms->method(), "sanity");
1062       // If the current throw can reach an exception handler in this JVMS,
1063       // then we must keep everything live that can reach that handler.
1064       // As a quick and dirty approximation, we look for any handlers at all.
1065       if (in_jvms->method()->has_exception_handlers()) {
1066         can_prune_locals = false;
1067       }
1068     }
1069 
1070     // Add the Locals
1071     k = in_jvms->locoff();
1072     l = in_jvms->loc_size();
1073     out_jvms->set_locoff(p);
1074     if (!can_prune_locals) {
1075       for (j = 0; j < l; j++) {
1076         call->set_req(p++, in_map->in(k + j));
1077       }
1078     } else {
1079       p += l;  // already set to top above by add_req_batch
1080     }
1081 
1082     // Add the Expression Stack
1083     k = in_jvms->stkoff();
1084     l = in_jvms->sp();
1085     out_jvms->set_stkoff(p);
1086     if (!can_prune_locals) {
1087       for (j = 0; j < l; j++) {
1088         call->set_req(p++, in_map->in(k + j));
1089       }
1090     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1091       // Divide stack into {S0,...,S1}, where S0 is set to top.
1092       uint s1 = stack_slots_not_pruned;
1093       stack_slots_not_pruned = 0;  // for next iteration
1094       if (s1 > l)  s1 = l;
1095       uint s0 = l - s1;
1096       p += s0;  // skip the tops preinstalled by add_req_batch
1097       for (j = s0; j < l; j++)
1098         call->set_req(p++, in_map->in(k+j));
1099     } else {
1100       p += l;  // already set to top above by add_req_batch
1101     }
1102 
1103     // Add the Monitors
1104     k = in_jvms->monoff();
1105     l = in_jvms->mon_size();
1106     out_jvms->set_monoff(p);
1107     for (j = 0; j < l; j++)
1108       call->set_req(p++, in_map->in(k+j));
1109 

1297   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1298   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1299   return _gvn.transform( new AndLNode(conv, mask) );
1300 }
1301 
1302 Node* GraphKit::ConvL2I(Node* offset) {
1303   // short-circuit a common case
1304   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1305   if (offset_con != (jlong)Type::OffsetBot) {
1306     return intcon((int) offset_con);
1307   }
1308   return _gvn.transform( new ConvL2INode(offset));
1309 }
1310 
1311 //-------------------------load_object_klass-----------------------------------
1312 Node* GraphKit::load_object_klass(Node* obj) {
1313   // Special-case a fresh allocation to avoid building nodes:
1314   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1315   if (akls != nullptr)  return akls;
1316   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1317   return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1318 }
1319 
1320 //-------------------------load_array_length-----------------------------------
1321 Node* GraphKit::load_array_length(Node* array) {
1322   // Special-case a fresh allocation to avoid building nodes:
1323   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1324   Node *alen;
1325   if (alloc == nullptr) {
1326     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1327     alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1328   } else {
1329     alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1330   }
1331   return alen;
1332 }
1333 
1334 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1335                                    const TypeOopPtr* oop_type,
1336                                    bool replace_length_in_map) {
1337   Node* length = alloc->Ideal_length();

1346         replace_in_map(length, ccast);
1347       }
1348       return ccast;
1349     }
1350   }
1351   return length;
1352 }
1353 
1354 //------------------------------do_null_check----------------------------------
1355 // Helper function to do a null pointer check.  Returned value is
1356 // the incoming address with null casted away.  You are allowed to use the
1357 // not-null value only if you are control dependent on the test.
1358 #ifndef PRODUCT
1359 extern uint explicit_null_checks_inserted,
1360             explicit_null_checks_elided;
1361 #endif
1362 Node* GraphKit::null_check_common(Node* value, BasicType type,
1363                                   // optional arguments for variations:
1364                                   bool assert_null,
1365                                   Node* *null_control,
1366                                   bool speculative,
1367                                   bool null_marker_check) {
1368   assert(!assert_null || null_control == nullptr, "not both at once");
1369   if (stopped())  return top();
1370   NOT_PRODUCT(explicit_null_checks_inserted++);
1371 
1372   if (value->is_InlineType()) {
1373     // Null checking a scalarized but nullable inline type. Check the null marker
1374     // input instead of the oop input to avoid keeping buffer allocations alive.
1375     null_check_common(value->as_InlineType()->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1376     if (stopped()) {
1377       return top();
1378     }
1379     if (assert_null) {
1380       // TODO 8350865 Scalarize here (this leads to failures with TestLWorld::test45)
1381       // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1382       // replace_in_map(value, vtptr);
1383       // return vtptr;
1384       replace_in_map(value, null());
1385       return null();
1386     }
1387     bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1388     return cast_not_null(value, do_replace_in_map);
1389   }
1390 
1391   // Construct null check
1392   Node *chk = nullptr;
1393   switch(type) {
1394     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1395     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1396     case T_ARRAY  : // fall through
1397       type = T_OBJECT;  // simplify further tests
1398     case T_OBJECT : {
1399       const Type *t = _gvn.type( value );
1400 
1401       const TypeOopPtr* tp = t->isa_oopptr();
1402       if (tp != nullptr && !tp->is_loaded()
1403           // Only for do_null_check, not any of its siblings:
1404           && !assert_null && null_control == nullptr) {
1405         // Usually, any field access or invocation on an unloaded oop type
1406         // will simply fail to link, since the statically linked class is
1407         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1408         // the static class is loaded but the sharper oop type is not.
1409         // Rather than checking for this obscure case in lots of places,
1410         // we simply observe that a null check on an unloaded class

1474         }
1475         Node *oldcontrol = control();
1476         set_control(cfg);
1477         Node *res = cast_not_null(value);
1478         set_control(oldcontrol);
1479         NOT_PRODUCT(explicit_null_checks_elided++);
1480         return res;
1481       }
1482       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1483       if (cfg == nullptr)  break;  // Quit at region nodes
1484       depth++;
1485     }
1486   }
1487 
1488   //-----------
1489   // Branch to failure if null
1490   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1491   Deoptimization::DeoptReason reason;
1492   if (assert_null) {
1493     reason = Deoptimization::reason_null_assert(speculative);
1494   } else if (type == T_OBJECT || null_marker_check) {
1495     reason = Deoptimization::reason_null_check(speculative);
1496   } else {
1497     reason = Deoptimization::Reason_div0_check;
1498   }
1499   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1500   // ciMethodData::has_trap_at will return a conservative -1 if any
1501   // must-be-null assertion has failed.  This could cause performance
1502   // problems for a method after its first do_null_assert failure.
1503   // Consider using 'Reason_class_check' instead?
1504 
1505   // To cause an implicit null check, we set the not-null probability
1506   // to the maximum (PROB_MAX).  For an explicit check the probability
1507   // is set to a smaller value.
1508   if (null_control != nullptr || too_many_traps(reason)) {
1509     // probability is less likely
1510     ok_prob =  PROB_LIKELY_MAG(3);
1511   } else if (!assert_null &&
1512              (ImplicitNullCheckThreshold > 0) &&
1513              method() != nullptr &&
1514              (method()->method_data()->trap_count(reason)

1548   }
1549 
1550   if (assert_null) {
1551     // Cast obj to null on this path.
1552     replace_in_map(value, zerocon(type));
1553     return zerocon(type);
1554   }
1555 
1556   // Cast obj to not-null on this path, if there is no null_control.
1557   // (If there is a null_control, a non-null value may come back to haunt us.)
1558   if (type == T_OBJECT) {
1559     Node* cast = cast_not_null(value, false);
1560     if (null_control == nullptr || (*null_control) == top())
1561       replace_in_map(value, cast);
1562     value = cast;
1563   }
1564 
1565   return value;
1566 }
1567 

1568 //------------------------------cast_not_null----------------------------------
1569 // Cast obj to not-null on this path
1570 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1571   const Type* t = _gvn.type(obj);
1572   const Type* t_not_null = t->join_speculative(TypePtr::NOTNULL);
1573   if (t == t_not_null) {
1574     return obj;
1575   }
1576 
1577   Node* cast = new CastPPNode(control(), obj, t_not_null);
1578   cast = _gvn.transform(cast);
1579 
1580   // Scan for instances of 'obj' in the current JVM mapping.
1581   // These instances are known to be not-null after the test.
1582   if (do_replace_in_map) {
1583     replace_in_map(obj, cast);
1584   }
1585   return cast;
1586 }
1587 
1588 // Sometimes in intrinsics, we implicitly know an object is not null
1589 // (there's no actual null check) so we can cast it to not null. In
1590 // the course of optimizations, the input to the cast can become null.
1591 // In that case that data path will die and we need the control path
1592 // to become dead as well to keep the graph consistent. So we have to
1593 // add a check for null for which one branch can't be taken. It uses
1594 // an OpaqueConstantBool node that will cause the check to be removed after loop
1595 // opts so the test goes away and the compiled code doesn't execute a
1596 // useless check.
1597 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1598   if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1599     return value;
1600   }
1601   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1602   Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1603   Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1604   IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1605   _gvn.set_type(iff, iff->Value(&_gvn));

1644 //=============================================================================
1645 //--------------------------------memory---------------------------------------
1646 Node* GraphKit::memory(uint alias_idx) {
1647   MergeMemNode* mem = merged_memory();
1648   Node* p = mem->memory_at(alias_idx);
1649   assert(p != mem->empty_memory(), "empty");
1650   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1651   return p;
1652 }
1653 
1654 //-----------------------------reset_memory------------------------------------
1655 Node* GraphKit::reset_memory() {
1656   Node* mem = map()->memory();
1657   // do not use this node for any more parsing!
1658   DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1659   return _gvn.transform( mem );
1660 }
1661 
1662 //------------------------------set_all_memory---------------------------------
1663 void GraphKit::set_all_memory(Node* newmem) {
1664   // The 2 cases are semantically equivalent
1665   MergeMemNode* mergemem;
1666   if (_gvn.is_IterGVN()) {
1667     // During IGVN, create a more predictable pattern so it is easier to verify that the GraphKit
1668     // does not modify memory
1669     mergemem = MergeMemNode::make(C->top());
1670     mergemem->set_base_memory(newmem);
1671   } else {
1672     // During parsing, be a little more aggressive so that GVN can fold accesses more easily
1673     mergemem = MergeMemNode::make(newmem);
1674   }
1675   _gvn.set_type_bottom(mergemem);
1676   record_for_igvn(mergemem);
1677   map()->set_memory(mergemem);
1678 }
1679 
1680 //------------------------------set_all_memory_call----------------------------
1681 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1682   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1683   set_all_memory(newmem);
1684 }
1685 
1686 //=============================================================================
1687 //
1688 // parser factory methods for MemNodes
1689 //
1690 // These are layered on top of the factory methods in LoadNode and StoreNode,
1691 // and integrate with the parser's memory state and _gvn engine.
1692 //
1693 
1694 // factory methods in "int adr_idx"
1695 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1696                           MemNode::MemOrd mo,
1697                           LoadNode::ControlDependency control_dependency,
1698                           bool require_atomic_access,
1699                           bool unaligned,
1700                           bool mismatched,
1701                           bool unsafe,
1702                           uint8_t barrier_data) {
1703   int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1704   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1705   const TypePtr* adr_type = nullptr; // debug-mode-only argument
1706   DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1707   Node* mem = memory(adr_idx);
1708   Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1709   ld = _gvn.transform(ld);
1710 
1711   if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1712     // Improve graph before escape analysis and boxing elimination.
1713     record_for_igvn(ld);
1714     if (ld->is_DecodeN()) {
1715       // Also record the actual load (LoadN) in case ld is DecodeN. In some
1716       // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1717       // a Phi). Recording such cases is still perfectly sound, but may be
1718       // unnecessary and result in some minor IGVN overhead.
1719       record_for_igvn(ld->in(1));
1720     }
1721   }
1722   return ld;
1723 }
1724 
1725 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1726                                 MemNode::MemOrd mo,
1727                                 bool require_atomic_access,
1728                                 bool unaligned,
1729                                 bool mismatched,
1730                                 bool unsafe,

1744   if (unsafe) {
1745     st->as_Store()->set_unsafe_access();
1746   }
1747   st->as_Store()->set_barrier_data(barrier_data);
1748   st = _gvn.transform(st);
1749   set_memory(st, adr_idx);
1750   // Back-to-back stores can only remove intermediate store with DU info
1751   // so push on worklist for optimizer.
1752   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1753     record_for_igvn(st);
1754 
1755   return st;
1756 }
1757 
1758 Node* GraphKit::access_store_at(Node* obj,
1759                                 Node* adr,
1760                                 const TypePtr* adr_type,
1761                                 Node* val,
1762                                 const Type* val_type,
1763                                 BasicType bt,
1764                                 DecoratorSet decorators,
1765                                 bool safe_for_replace,
1766                                 const InlineTypeNode* vt) {
1767   // Transformation of a value which could be null pointer (CastPP #null)
1768   // could be delayed during Parse (for example, in adjust_map_after_if()).
1769   // Execute transformation here to avoid barrier generation in such case.
1770   if (_gvn.type(val) == TypePtr::NULL_PTR) {
1771     val = _gvn.makecon(TypePtr::NULL_PTR);
1772   }
1773 
1774   if (stopped()) {
1775     return top(); // Dead path ?
1776   }
1777 
1778   assert(val != nullptr, "not dead path");
1779   if (val->is_InlineType()) {
1780     // Store to non-flat field. Buffer the inline type and make sure
1781     // the store is re-executed if the allocation triggers deoptimization.
1782     PreserveReexecuteState preexecs(this);
1783     jvms()->set_should_reexecute(true);
1784     val = val->as_InlineType()->buffer(this, safe_for_replace);
1785   }
1786 
1787   C2AccessValuePtr addr(adr, adr_type);
1788   C2AccessValue value(val, val_type);
1789   C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1790   if (access.is_raw()) {
1791     return _barrier_set->BarrierSetC2::store_at(access, value);
1792   } else {
1793     return _barrier_set->store_at(access, value);
1794   }
1795 }
1796 
1797 Node* GraphKit::access_load_at(Node* obj,   // containing obj
1798                                Node* adr,   // actual address to store val at
1799                                const TypePtr* adr_type,
1800                                const Type* val_type,
1801                                BasicType bt,
1802                                DecoratorSet decorators,
1803                                Node* ctl) {
1804   if (stopped()) {
1805     return top(); // Dead path ?
1806   }
1807 
1808   SavedState old_state(this);
1809   C2AccessValuePtr addr(adr, adr_type);
1810   C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1811   Node* load;
1812   if (access.is_raw()) {
1813     load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1814   } else {
1815     load = _barrier_set->load_at(access, val_type);
1816   }
1817 
1818   // Restore the previous state only if the load got folded to a constant
1819   // and we can discard any barriers that might have been added.
1820   if (load == nullptr || !load->is_Con()) {
1821     old_state.discard();
1822   }
1823   return load;
1824 }
1825 
1826 Node* GraphKit::access_load(Node* adr,   // actual address to load val at
1827                             const Type* val_type,
1828                             BasicType bt,
1829                             DecoratorSet decorators) {
1830   if (stopped()) {

1912                                      Node* new_val,
1913                                      const Type* value_type,
1914                                      BasicType bt,
1915                                      DecoratorSet decorators) {
1916   C2AccessValuePtr addr(adr, adr_type);
1917   C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1918   if (access.is_raw()) {
1919     return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1920   } else {
1921     return _barrier_set->atomic_add_at(access, new_val, value_type);
1922   }
1923 }
1924 
1925 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1926   return _barrier_set->clone(this, src, dst, size, is_array);
1927 }
1928 
1929 //-------------------------array_element_address-------------------------
1930 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1931                                       const TypeInt* sizetype, Node* ctrl) {
1932   const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1933   uint shift;
1934   uint header;
1935   if (arytype->is_flat() && arytype->klass_is_exact()) {
1936     // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1937     // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1938     // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1939     // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1940     // though we don't need the address node in this case and throw it away again.
1941     shift = arytype->flat_log_elem_size();
1942     header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1943   } else {
1944     shift = exact_log2(type2aelembytes(elembt));
1945     header = arrayOopDesc::base_offset_in_bytes(elembt);
1946   }
1947 
1948   // short-circuit a common case (saves lots of confusing waste motion)
1949   jint idx_con = find_int_con(idx, -1);
1950   if (idx_con >= 0) {
1951     intptr_t offset = header + ((intptr_t)idx_con << shift);
1952     return basic_plus_adr(ary, offset);
1953   }
1954 
1955   // must be correct type for alignment purposes
1956   Node* base  = basic_plus_adr(ary, header);
1957   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1958   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1959   return basic_plus_adr(ary, base, scale);
1960 }
1961 
1962 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1963   assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1964   if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1965     return cast_to_flat_array_exact(array, elem_vk, true, false);
1966   } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1967     return cast_to_flat_array_exact(array, elem_vk, true, true);
1968   } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1969     return cast_to_flat_array_exact(array, elem_vk, false, true);
1970   }
1971 
1972   bool is_null_free = false;
1973   if (!elem_vk->has_nullable_atomic_layout()) {
1974     // Element does not have a nullable flat layout, cannot be nullable
1975     is_null_free = true;
1976   }
1977 
1978   ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1979   const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1980   arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1981   return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1982 }
1983 
1984 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1985   assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1986   ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1987   const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1988   assert(arytype->klass_is_exact(), "inconsistency");
1989   assert(arytype->is_flat(), "inconsistency");
1990   assert(arytype->is_null_free() == is_null_free, "inconsistency");
1991   assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1992   return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1993 }
1994 
1995 //-------------------------load_array_element-------------------------
1996 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1997   const Type* elemtype = arytype->elem();
1998   BasicType elembt = elemtype->array_element_basic_type();
1999   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
2000   if (elembt == T_NARROWOOP) {
2001     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
2002   }
2003   Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
2004                             IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
2005   return ld;
2006 }
2007 
2008 //-------------------------set_arguments_for_java_call-------------------------
2009 // Arguments (pre-popped from the stack) are taken from the JVMS.
2010 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
2011   PreserveReexecuteState preexecs(this);
2012   if (Arguments::is_valhalla_enabled()) {
2013     // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
2014     // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
2015     jvms()->set_should_reexecute(true);
2016     int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
2017     inc_sp(arg_size);
2018   }
2019   // Add the call arguments
2020   const TypeTuple* domain = call->tf()->domain_sig();
2021   uint nargs = domain->cnt();
2022   int arg_num = 0;
2023   for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
2024     uint arg_idx = i - TypeFunc::Parms;
2025     Node* arg = argument(arg_idx);
2026     const Type* t = domain->field_at(i);
2027     if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
2028       // We don't pass inline type arguments by reference but instead pass each field of the inline type
2029       if (!arg->is_InlineType()) {
2030         // There are 2 cases in which the argument has not been scalarized
2031         if (_gvn.type(arg)->is_zero_type()) {
2032           arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
2033         } else {
2034           // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
2035           // receiver is a non-scalarized oop. CHA or IGVN might then prove that the receiver
2036           // type must be an exact value class. The method is devirtualized, and replaced with
2037           // a direct call with a scalarized receiver instead.
2038           assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
2039           assert(call->is_optimized_virtual(), "must be during devirtualization of calls");
2040           arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
2041         }
2042       }
2043       InlineTypeNode* vt = arg->as_InlineType();
2044       vt->pass_fields(this, call, idx, true, !t->maybe_null(), true);
2045       // If an inline type argument is passed as fields, attach the Method* to the call site
2046       // to be able to access the extended signature later via attached_method_before_pc().
2047       // For example, see CompiledMethod::preserve_callee_argument_oops().
2048       call->set_override_symbolic_info(true);
2049       // Register a calling convention dependency on the callee method to make sure that this method is deoptimized and
2050       // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
2051       C->dependencies()->assert_mismatch_calling_convention(call->method());
2052       arg_num++;
2053       continue;
2054     } else if (arg->is_InlineType()) {
2055       // Pass inline type argument via oop to callee
2056       arg = arg->as_InlineType()->buffer(this, true);
2057     }
2058     if (t != Type::HALF) {
2059       arg_num++;
2060     }
2061     call->init_req(idx++, arg);
2062   }
2063 }
2064 
2065 //---------------------------set_edges_for_java_call---------------------------
2066 // Connect a newly created call into the current JVMS.
2067 // A return value node (if any) is returned from set_edges_for_java_call.
2068 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2069 
2070   // Add the predefined inputs:
2071   call->init_req( TypeFunc::Control, control() );
2072   call->init_req( TypeFunc::I_O    , i_o() );
2073   call->init_req( TypeFunc::Memory , reset_memory() );
2074   call->init_req( TypeFunc::FramePtr, frameptr() );
2075   call->init_req( TypeFunc::ReturnAdr, top() );
2076 
2077   add_safepoint_edges(call, must_throw);
2078 
2079   Node* xcall = _gvn.transform(call);
2080 
2081   if (xcall == top()) {
2082     set_control(top());
2083     return;
2084   }
2085   assert(xcall == call, "call identity is stable");
2086 
2087   // Re-use the current map to produce the result.
2088 
2089   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2090   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
2091   set_all_memory_call(xcall, separate_io_proj);
2092 
2093   //return xcall;   // no need, caller already has it
2094 }
2095 
2096 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2097   if (stopped())  return top();  // maybe the call folded up?
2098 







2099   // Note:  Since any out-of-line call can produce an exception,
2100   // we always insert an I_O projection from the call into the result.
2101 
2102   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2103 
2104   if (separate_io_proj) {
2105     // The caller requested separate projections be used by the fall
2106     // through and exceptional paths, so replace the projections for
2107     // the fall through path.
2108     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2109     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2110   }
2111 
2112   // Capture the return value, if any.
2113   Node* ret;
2114   if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2115     ret = top();
2116   } else if (call->tf()->returns_inline_type_as_fields()) {
2117     // Return of multiple values (inline type fields): we create a
2118     // InlineType node, each field is a projection from the call.
2119     ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2120     uint base_input = TypeFunc::Parms;
2121     ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2122     // If we run out of registers to store the null marker, we need to reserve an extra
2123     // slot to store it on the stack. Unfortunately, we only know if stack slot is needed
2124     // when matching the call (see Matcher::return_values_mask), so we are conservative here.
2125     C->set_needs_nm_slot(true);
2126   } else {
2127     ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2128     ciType* t = call->method()->return_type();
2129     if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2130       // The return type is unloaded but the callee might later be C2 compiled and then return
2131       // in scalarized form when the return type is loaded. Handle this similar to what we do in
2132       // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2133       // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2134       IdealKit ideal(this);
2135       IdealVariable res(ideal);
2136       ideal.declarations_done();
2137       // Change return type of call to scalarized return
2138       const TypeFunc* tf = call->_tf;
2139       const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2140       const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2141       call->_tf = new_tf;
2142       _gvn.set_type(call, call->Value(&_gvn));
2143       _gvn.set_type(ret, ret->Value(&_gvn));
2144       // Don't add store to buffer call if we are strength reducing
2145       if (!C->strength_reduction()) {
2146         ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2147           // Return value is null
2148           ideal.set(res, makecon(TypePtr::NULL_PTR));
2149         } ideal.else_(); {
2150           // Return value is non-null
2151           sync_kit(ideal);
2152 
2153           Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2154                                                       OptoRuntime::store_inline_type_fields_Type(),
2155                                                       StubRoutines::store_inline_type_fields_to_buf(),
2156                                                       nullptr, TypePtr::BOTTOM, ret);
2157 
2158           // We don't know how many values are returned. This assumes the
2159           // worst case, that all available registers are used.
2160           for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2161             if (domain->field_at(i) == Type::HALF) {
2162               store_to_buf_call->init_req(i, top());
2163               continue;
2164             }
2165             Node* proj =_gvn.transform(new ProjNode(call, i));
2166             store_to_buf_call->init_req(i, proj);
2167           }
2168           make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2169 
2170           Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2171           const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2172           buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2173 
2174           ideal.set(res, buf);
2175           ideal.sync_kit(this);
2176         } ideal.end_if();
2177       } else {
2178         for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2179           // Will be rewired later in replace_call().
2180           _gvn.transform(new ProjNode(call, i));
2181         }
2182         ideal.set(res, ret);
2183       }
2184       sync_kit(ideal);
2185       ret = _gvn.transform(ideal.value(res));
2186     } else if (!call->method()->return_value_is_larval() && _gvn.type(ret)->is_inlinetypeptr()) {
2187       // In Parse::do_call we call make_from_oop on the final result of the call, but this could be the
2188       // result of merging several call paths. If one of them is made of an actual call node that
2189       // returns an oop, we need to call make_from_oop here as well because we want InlineType
2190       // nodes on every path to avoid merging an unallocated InlineType node path with an oop path.
2191       ret = InlineTypeNode::make_from_oop(this, ret, _gvn.type(ret)->inline_klass());
2192     }
2193   }
2194 
2195   return ret;
2196 }
2197 
2198 //--------------------set_predefined_input_for_runtime_call--------------------
2199 // Reading and setting the memory state is way conservative here.
2200 // The real problem is that I am not doing real Type analysis on memory,
2201 // so I cannot distinguish card mark stores from other stores.  Across a GC
2202 // point the Store Barrier and the card mark memory has to agree.  I cannot
2203 // have a card mark store and its barrier split across the GC point from
2204 // either above or below.  Here I get that to happen by reading ALL of memory.
2205 // A better answer would be to separate out card marks from other memory.
2206 // For now, return the input memory state, so that it can be reused
2207 // after the call, if this call has restricted memory effects.
2208 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2209   // Set fixed predefined input arguments
2210   call->init_req(TypeFunc::Control, control());
2211   call->init_req(TypeFunc::I_O, top()); // does no i/o
2212   call->init_req(TypeFunc::ReturnAdr, top());
2213   if (call->is_CallLeafPure()) {
2214     call->init_req(TypeFunc::Memory, top());

2276     if (use->is_MergeMem()) {
2277       wl.push(use);
2278     }
2279   }
2280 }
2281 
2282 // Replace the call with the current state of the kit.
2283 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2284   JVMState* ejvms = nullptr;
2285   if (has_exceptions()) {
2286     ejvms = transfer_exceptions_into_jvms();
2287   }
2288 
2289   ReplacedNodes replaced_nodes = map()->replaced_nodes();
2290   ReplacedNodes replaced_nodes_exception;
2291   Node* ex_ctl = top();
2292 
2293   SafePointNode* final_state = stop();
2294 
2295   // Find all the needed outputs of this call
2296   CallProjections* callprojs = call->extract_projections(true, do_asserts);

2297 
2298   Unique_Node_List wl;

2299   Node* final_mem = final_state->in(TypeFunc::Memory);
2300   Node* final_ctl = final_state->in(TypeFunc::Control);
2301   Node* final_io = final_state->in(TypeFunc::I_O);
2302 
2303   // Replace all the old call edges with the edges from the inlining result
2304   if (callprojs->fallthrough_catchproj != nullptr) {
2305     C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2306   }
2307   if (callprojs->fallthrough_memproj != nullptr) {
2308     if (final_mem->is_MergeMem()) {
2309       // Parser's exits MergeMem was not transformed but may be optimized
2310       final_mem = _gvn.transform(final_mem);
2311     }
2312     C->gvn_replace_by(callprojs->fallthrough_memproj,   final_mem);
2313     add_mergemem_users_to_worklist(wl, final_mem);
2314   }
2315   if (callprojs->fallthrough_ioproj != nullptr) {
2316     C->gvn_replace_by(callprojs->fallthrough_ioproj,    final_io);
2317   }
2318 
2319   // Replace the result with the new result if it exists and is used
2320   if (callprojs->resproj[0] != nullptr && result != nullptr) {
2321     // If the inlined code is dead, the result projections for an inline type returned as
2322     // fields have not been replaced. They will go away once the call is replaced by TOP below.
2323     assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2324            (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2325            "unexpected number of results");
2326     // If we are doing strength reduction and the return type is not loaded we
2327     // need to rewire all projections since store_inline_type_fields_to_buf is already present
2328     if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2329       CallNode* new_call = result->in(0)->as_Call();
2330       assert(new_call->proj_out_or_null(TypeFunc::Parms) == result, "the first data projection should be result");
2331       for (uint i = 0; i < callprojs->nb_resproj; i++) {
2332         if (callprojs->resproj[i] != nullptr) {
2333           Node* new_proj = new_call->proj_out_or_null(TypeFunc::Parms + i);
2334           assert(new_proj != nullptr, "projection should be available");
2335           C->gvn_replace_by(callprojs->resproj[i], new_proj);
2336         }
2337       }
2338     } else {
2339       C->gvn_replace_by(callprojs->resproj[0], result);
2340     }
2341   }
2342 
2343   if (ejvms == nullptr) {
2344     // No exception edges to simply kill off those paths
2345     if (callprojs->catchall_catchproj != nullptr) {
2346       C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2347     }
2348     if (callprojs->catchall_memproj != nullptr) {
2349       C->gvn_replace_by(callprojs->catchall_memproj,   C->top());
2350     }
2351     if (callprojs->catchall_ioproj != nullptr) {
2352       C->gvn_replace_by(callprojs->catchall_ioproj,    C->top());
2353     }
2354     // Replace the old exception object with top
2355     if (callprojs->exobj != nullptr) {
2356       C->gvn_replace_by(callprojs->exobj, C->top());
2357     }
2358   } else {
2359     GraphKit ekit(ejvms);
2360 
2361     // Load my combined exception state into the kit, with all phis transformed:
2362     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2363     replaced_nodes_exception = ex_map->replaced_nodes();
2364 
2365     Node* ex_oop = ekit.use_exception_state(ex_map);
2366 
2367     if (callprojs->catchall_catchproj != nullptr) {
2368       C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2369       ex_ctl = ekit.control();
2370     }
2371     if (callprojs->catchall_memproj != nullptr) {
2372       Node* ex_mem = ekit.reset_memory();
2373       C->gvn_replace_by(callprojs->catchall_memproj,   ex_mem);
2374       add_mergemem_users_to_worklist(wl, ex_mem);
2375     }
2376     if (callprojs->catchall_ioproj != nullptr) {
2377       C->gvn_replace_by(callprojs->catchall_ioproj,    ekit.i_o());
2378     }
2379 
2380     // Replace the old exception object with the newly created one
2381     if (callprojs->exobj != nullptr) {
2382       C->gvn_replace_by(callprojs->exobj, ex_oop);
2383     }
2384   }
2385 
2386   // Disconnect the call from the graph
2387   call->disconnect_inputs(C);
2388   C->gvn_replace_by(call, C->top());
2389 
2390   // Clean up any MergeMems that feed other MergeMems since the
2391   // optimizer doesn't like that.
2392   while (wl.size() > 0) {
2393     _gvn.transform(wl.pop());
2394   }
2395 
2396   if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2397     replaced_nodes.apply(C, final_ctl);
2398   }
2399   if (!ex_ctl->is_top() && do_replaced_nodes) {
2400     replaced_nodes_exception.apply(C, ex_ctl);
2401   }
2402 }
2403 
2404 
2405 //------------------------------increment_counter------------------------------
2406 // for statistics: increment a VM counter by 1
2407 
2408 void GraphKit::increment_counter(address counter_addr) {
2409   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2410   increment_counter(adr1);
2411 }
2412 
2413 void GraphKit::increment_counter(Node* counter_addr) {
2414   Node* ctrl = control();
2415   Node* cnt  = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2416   Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2417   store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2418 }
2419 
2420 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2421   Node* halt = new HaltNode(ctrl, frameptr, reason
2422                             PRODUCT_ONLY(COMMA generate_code_in_product));
2423   halt = _gvn.transform(halt);
2424   root()->add_req(halt);
2425   if (_gvn.is_IterGVN() != nullptr) {
2426     record_for_igvn(root());
2427   }
2428 }
2429 
2430 //------------------------------uncommon_trap----------------------------------
2431 // Bail out to the interpreter in mid-method.  Implemented by calling the
2432 // uncommon_trap blob.  This helper function inserts a runtime call with the
2433 // right debug info.
2434 Node* GraphKit::uncommon_trap(int trap_request,
2435                              ciKlass* klass, const char* comment,
2436                              bool must_throw,
2437                              bool keep_exact_action) {
2438   if (failing_internal()) {
2439     stop();
2440   }
2441   if (stopped())  return nullptr; // trap reachable?
2442 
2443   // Note:  If ProfileTraps is true, and if a deopt. actually
2444   // occurs here, the runtime will make sure an MDO exists.  There is
2445   // no need to call method()->ensure_method_data() at this point.
2446 
2447   // Set the stack pointer to the right value for reexecution:

2589  *
2590  * @param n          node that the type applies to
2591  * @param exact_kls  type from profiling
2592  * @param maybe_null did profiling see null?
2593  *
2594  * @return           node with improved type
2595  */
2596 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2597   const Type* current_type = _gvn.type(n);
2598   assert(UseTypeSpeculation, "type speculation must be on");
2599 
2600   const TypePtr* speculative = current_type->speculative();
2601 
2602   // Should the klass from the profile be recorded in the speculative type?
2603   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2604     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2605     const TypeOopPtr* xtype = tklass->as_instance_type();
2606     assert(xtype->klass_is_exact(), "Should be exact");
2607     // Any reason to believe n is not null (from this profiling or a previous one)?
2608     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2609     const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2610     // record the new speculative type's depth
2611     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2612     speculative = speculative->with_inline_depth(jvms()->depth());
2613   } else if (current_type->would_improve_ptr(ptr_kind)) {
2614     // Profiling report that null was never seen so we can change the
2615     // speculative type to non null ptr.
2616     if (ptr_kind == ProfileAlwaysNull) {
2617       speculative = TypePtr::NULL_PTR;
2618     } else {
2619       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2620       const TypePtr* ptr = TypePtr::NOTNULL;
2621       if (speculative != nullptr) {
2622         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2623       } else {
2624         speculative = ptr;
2625       }
2626     }
2627   }
2628 
2629   if (speculative != current_type->speculative()) {
2630     // Build a type with a speculative type (what we think we know
2631     // about the type but will need a guard when we use it)
2632     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2633     // We're changing the type, we need a new CheckCast node to carry
2634     // the new type. The new type depends on the control: what
2635     // profiling tells us is only valid from here as far as we can
2636     // tell.
2637     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2638     cast = _gvn.transform(cast);
2639     replace_in_map(n, cast);
2640     n = cast;
2641   }
2642 
2643   return n;
2644 }
2645 
2646 /**
2647  * Record profiling data from receiver profiling at an invoke with the
2648  * type system so that it can propagate it (speculation)
2649  *
2650  * @param n  receiver node
2651  *
2652  * @return   node with improved type
2653  */
2654 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2655   if (!UseTypeSpeculation) {
2656     return n;
2657   }
2658   ciKlass* exact_kls = profile_has_unique_klass();
2659   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2660   if ((java_bc() == Bytecodes::_checkcast ||
2661        java_bc() == Bytecodes::_instanceof ||
2662        java_bc() == Bytecodes::_aastore) &&
2663       method()->method_data()->is_mature()) {
2664     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2665     if (data != nullptr) {
2666       if (java_bc() == Bytecodes::_aastore) {
2667         ciKlass* array_type = nullptr;
2668         ciKlass* element_type = nullptr;
2669         ProfilePtrKind element_ptr = ProfileMaybeNull;
2670         bool flat_array = true;
2671         bool null_free_array = true;
2672         method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2673         exact_kls = element_type;
2674         ptr_kind = element_ptr;
2675       } else {
2676         if (!data->as_BitData()->null_seen()) {
2677           ptr_kind = ProfileNeverNull;
2678         } else {
2679           if (TypeProfileCasts) {
2680             assert(data->is_ReceiverTypeData(), "bad profile data type");
2681             ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2682             uint i = 0;
2683             for (; i < call->row_limit(); i++) {
2684               ciKlass* receiver = call->receiver(i);
2685               if (receiver != nullptr) {
2686                 break;
2687               }
2688             }
2689             ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2690           }

2691         }
2692       }
2693     }
2694   }
2695   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2696 }
2697 
2698 /**
2699  * Record profiling data from argument profiling at an invoke with the
2700  * type system so that it can propagate it (speculation)
2701  *
2702  * @param dest_method  target method for the call
2703  * @param bc           what invoke bytecode is this?
2704  */
2705 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2706   if (!UseTypeSpeculation) {
2707     return;
2708   }
2709   const TypeFunc* tf    = TypeFunc::make(dest_method);
2710   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2711   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2712   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2713     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2714     if (is_reference_type(targ->basic_type())) {
2715       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2716       ciKlass* better_type = nullptr;
2717       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2718         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2719       }
2720       i++;
2721     }
2722   }
2723 }
2724 
2725 /**
2726  * Record profiling data from parameter profiling at an invoke with
2727  * the type system so that it can propagate it (speculation)
2728  */
2729 void GraphKit::record_profiled_parameters_for_speculation() {
2730   if (!UseTypeSpeculation) {
2731     return;
2732   }
2733   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {

2853                                   // The first null ends the list.
2854                                   Node* parm0, Node* parm1,
2855                                   Node* parm2, Node* parm3,
2856                                   Node* parm4, Node* parm5,
2857                                   Node* parm6, Node* parm7) {
2858   assert(call_addr != nullptr, "must not call null targets");
2859 
2860   // Slow-path call
2861   bool is_leaf = !(flags & RC_NO_LEAF);
2862   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2863   if (call_name == nullptr) {
2864     assert(!is_leaf, "must supply name for leaf");
2865     call_name = OptoRuntime::stub_name(call_addr);
2866   }
2867   CallNode* call;
2868   if (!is_leaf) {
2869     call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2870   } else if (flags & RC_NO_FP) {
2871     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2872   } else  if (flags & RC_VECTOR){
2873     uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2874     call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2875   } else if (flags & RC_PURE) {
2876     assert(adr_type == nullptr, "pure call does not touch memory");
2877     call = new CallLeafPureNode(call_type, call_addr, call_name);
2878   } else {
2879     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2880   }
2881 
2882   // The following is similar to set_edges_for_java_call,
2883   // except that the memory effects of the call are restricted to AliasIdxRaw.
2884 
2885   // Slow path call has no side-effects, uses few values
2886   bool wide_in  = !(flags & RC_NARROW_MEM);
2887   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2888 
2889   Node* prev_mem = nullptr;
2890   if (wide_in) {
2891     prev_mem = set_predefined_input_for_runtime_call(call);
2892   } else {
2893     assert(!wide_out, "narrow in => narrow out");
2894     Node* narrow_mem = memory(adr_type);
2895     prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2896   }
2897 
2898   // Hook each parm in order.  Stop looking at the first null.
2899   if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2900   if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2901   if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2902   if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2903   if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2904   if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2905   if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2906   if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2907   /* close each nested if ===> */  } } } } } } } }
2908   assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2909 
2910   if (!is_leaf) {
2911     // Non-leaves can block and take safepoints:
2912     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2913   }
2914   // Non-leaves can throw exceptions:
2915   if (has_io) {
2916     call->set_req(TypeFunc::I_O, i_o());
2917   }
2918 
2919   if (flags & RC_UNCOMMON) {
2920     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2921     // (An "if" probability corresponds roughly to an unconditional count.
2922     // Sort of.)
2923     call->set_cnt(PROB_UNLIKELY_MAG(4));
2924   }
2925 
2926   Node* c = _gvn.transform(call);
2927   assert(c == call, "cannot disappear");
2928 

2936 
2937   if (has_io) {
2938     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2939   }
2940   return call;
2941 
2942 }
2943 
2944 // i2b
2945 Node* GraphKit::sign_extend_byte(Node* in) {
2946   Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2947   return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2948 }
2949 
2950 // i2s
2951 Node* GraphKit::sign_extend_short(Node* in) {
2952   Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2953   return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2954 }
2955 
2956 
2957 //------------------------------merge_memory-----------------------------------
2958 // Merge memory from one path into the current memory state.
2959 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2960   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2961     Node* old_slice = mms.force_memory();
2962     Node* new_slice = mms.memory2();
2963     if (old_slice != new_slice) {
2964       PhiNode* phi;
2965       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2966         if (mms.is_empty()) {
2967           // clone base memory Phi's inputs for this memory slice
2968           assert(old_slice == mms.base_memory(), "sanity");
2969           phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2970           _gvn.set_type(phi, Type::MEMORY);
2971           for (uint i = 1; i < phi->req(); i++) {
2972             phi->init_req(i, old_slice->in(i));
2973           }
2974         } else {
2975           phi = old_slice->as_Phi(); // Phi was generated already
2976         }

3033   gvn.transform(iff);
3034   if (!bol->is_Con()) gvn.record_for_igvn(iff);
3035   return iff;
3036 }
3037 
3038 //-------------------------------gen_subtype_check-----------------------------
3039 // Generate a subtyping check.  Takes as input the subtype and supertype.
3040 // Returns 2 values: sets the default control() to the true path and returns
3041 // the false path.  Only reads invariant memory; sets no (visible) memory.
3042 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
3043 // but that's not exposed to the optimizer.  This call also doesn't take in an
3044 // Object; if you wish to check an Object you need to load the Object's class
3045 // prior to coming here.
3046 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
3047                                ciMethod* method, int bci) {
3048   Compile* C = gvn.C;
3049   if ((*ctrl)->is_top()) {
3050     return C->top();
3051   }
3052 
3053   const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
3054   // For a direct pointer comparison, we need the refined array klass pointer
3055   Node* vm_superklass = superklass;
3056   if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
3057     assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
3058     vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
3059   }
3060 
3061   // Fast check for identical types, perhaps identical constants.
3062   // The types can even be identical non-constants, in cases
3063   // involving Array.newInstance, Object.clone, etc.
3064   if (subklass == superklass)
3065     return C->top();             // false path is dead; no test needed.
3066 
3067   if (gvn.type(superklass)->singleton()) {
3068     const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3069     const TypeKlassPtr* subk   = gvn.type(subklass)->is_klassptr();
3070 
3071     // In the common case of an exact superklass, try to fold up the
3072     // test before generating code.  You may ask, why not just generate
3073     // the code and then let it fold up?  The answer is that the generated
3074     // code will necessarily include null checks, which do not always
3075     // completely fold away.  If they are also needless, then they turn
3076     // into a performance loss.  Example:
3077     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3078     // Here, the type of 'fa' is often exact, so the store check
3079     // of fa[1]=x will fold up, without testing the nullness of x.
3080     //
3081     // At macro expansion, we would have already folded the SubTypeCheckNode
3082     // being expanded here because we always perform the static sub type
3083     // check in SubTypeCheckNode::sub() regardless of whether
3084     // StressReflectiveCode is set or not. We can therefore skip this
3085     // static check when StressReflectiveCode is on.
3086     switch (C->static_subtype_check(superk, subk)) {
3087     case Compile::SSC_always_false:
3088       {
3089         Node* always_fail = *ctrl;
3090         *ctrl = gvn.C->top();
3091         return always_fail;
3092       }
3093     case Compile::SSC_always_true:
3094       return C->top();
3095     case Compile::SSC_easy_test:
3096       {
3097         // Just do a direct pointer compare and be done.
3098         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3099         *ctrl = gvn.transform(new IfTrueNode(iff));
3100         return gvn.transform(new IfFalseNode(iff));
3101       }
3102     case Compile::SSC_full_test:
3103       break;
3104     default:
3105       ShouldNotReachHere();
3106     }
3107   }
3108 
3109   // %%% Possible further optimization:  Even if the superklass is not exact,
3110   // if the subklass is the unique subtype of the superklass, the check
3111   // will always succeed.  We could leave a dependency behind to ensure this.
3112 
3113   // First load the super-klass's check-offset
3114   Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3115   Node* m = C->immutable_memory();
3116   Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3117   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3118   const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();

3156   gvn.record_for_igvn(r_ok_subtype);
3157 
3158   // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3159   // SubTypeCheck node
3160   if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3161     ciCallProfile profile = method->call_profile_at_bci(bci);
3162     float total_prob = 0;
3163     for (int i = 0; profile.has_receiver(i); ++i) {
3164       float prob = profile.receiver_prob(i);
3165       total_prob += prob;
3166     }
3167     if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3168       const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3169       for (int i = 0; profile.has_receiver(i); ++i) {
3170         ciKlass* klass = profile.receiver(i);
3171         const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3172         Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3173         if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3174           continue;
3175         }
3176         if (klass_t->isa_aryklassptr()) {
3177           // For a direct pointer comparison, we need the refined array klass pointer
3178           klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3179         }
3180         float prob = profile.receiver_prob(i);
3181         ConNode* klass_node = gvn.makecon(klass_t);
3182         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3183         Node* iftrue = gvn.transform(new IfTrueNode(iff));
3184 
3185         if (result == Compile::SSC_always_true) {
3186           r_ok_subtype->add_req(iftrue);
3187         } else {
3188           assert(result == Compile::SSC_always_false, "");
3189           r_not_subtype->add_req(iftrue);
3190         }
3191         *ctrl = gvn.transform(new IfFalseNode(iff));
3192       }
3193     }
3194   }
3195 
3196   // See if we get an immediate positive hit.  Happens roughly 83% of the
3197   // time.  Test to see if the value loaded just previously from the subklass
3198   // is exactly the superklass.
3199   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);

3213       igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3214     }
3215     return not_subtype_ctrl;
3216   }
3217 
3218   r_ok_subtype->init_req(1, iftrue1);
3219 
3220   // Check for immediate negative hit.  Happens roughly 11% of the time (which
3221   // is roughly 63% of the remaining cases).  Test to see if the loaded
3222   // check-offset points into the subklass display list or the 1-element
3223   // cache.  If it points to the display (and NOT the cache) and the display
3224   // missed then it's not a subtype.
3225   Node *cacheoff = gvn.intcon(cacheoff_con);
3226   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3227   r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3228   *ctrl = gvn.transform(new IfFalseNode(iff2));
3229 
3230   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
3231   // No performance impact (too rare) but allows sharing of secondary arrays
3232   // which has some footprint reduction.
3233   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3234   r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3235   *ctrl = gvn.transform(new IfFalseNode(iff3));
3236 
3237   // -- Roads not taken here: --
3238   // We could also have chosen to perform the self-check at the beginning
3239   // of this code sequence, as the assembler does.  This would not pay off
3240   // the same way, since the optimizer, unlike the assembler, can perform
3241   // static type analysis to fold away many successful self-checks.
3242   // Non-foldable self checks work better here in second position, because
3243   // the initial primary superclass check subsumes a self-check for most
3244   // types.  An exception would be a secondary type like array-of-interface,
3245   // which does not appear in its own primary supertype display.
3246   // Finally, we could have chosen to move the self-check into the
3247   // PartialSubtypeCheckNode, and from there out-of-line in a platform
3248   // dependent manner.  But it is worthwhile to have the check here,
3249   // where it can be perhaps be optimized.  The cost in code space is
3250   // small (register compare, branch).
3251 
3252   // Now do a linear scan of the secondary super-klass array.  Again, no real
3253   // performance impact (too rare) but it's gotta be done.
3254   // Since the code is rarely used, there is no penalty for moving it
3255   // out of line, and it can only improve I-cache density.
3256   // The decision to inline or out-of-line this final check is platform
3257   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3258   Node* psc = gvn.transform(
3259     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3260 
3261   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3262   r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3263   r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3264 
3265   // Return false path; set default control to true path.
3266   *ctrl = gvn.transform(r_ok_subtype);
3267   return gvn.transform(r_not_subtype);
3268 }
3269 
3270 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3271   const Type* sub_t = _gvn.type(obj_or_subklass);
3272   if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3273     sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3274     obj_or_subklass = makecon(sub_t);
3275   }
3276   bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3277   if (expand_subtype_check) {
3278     MergeMemNode* mem = merged_memory();
3279     Node* ctrl = control();
3280     Node* subklass = obj_or_subklass;
3281     if (!sub_t->isa_klassptr()) {
3282       subklass = load_object_klass(obj_or_subklass);
3283     }
3284 
3285     Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3286     set_control(ctrl);
3287     return n;
3288   }
3289 
3290   Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3291   Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3292   IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3293   set_control(_gvn.transform(new IfTrueNode(iff)));
3294   return _gvn.transform(new IfFalseNode(iff));
3295 }
3296 
3297 // Profile-driven exact type check:
3298 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3299                                     float prob, Node* *casted_receiver) {

3300   assert(!klass->is_interface(), "no exact type check on interfaces");
3301   Node* fail = top();
3302   const Type* rec_t = _gvn.type(receiver);
3303   if (rec_t->is_inlinetypeptr()) {
3304     if (klass->equals(rec_t->inline_klass())) {
3305       (*casted_receiver) = receiver; // Always passes
3306     } else {
3307       (*casted_receiver) = top();    // Always fails
3308       fail = control();
3309       set_control(top());
3310     }
3311     return fail;
3312   }
3313   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3314   if (tklass->isa_aryklassptr()) {
3315     // For a direct pointer comparison, we need the refined array klass pointer
3316     tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3317   }
3318   Node* recv_klass = load_object_klass(receiver);
3319   fail = type_check(recv_klass, tklass, prob);





3320 
3321   if (!stopped()) {
3322     const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3323     const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3324     assert(recv_xtype->klass_is_exact(), "");
3325 
3326     if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3327       // Subsume downstream occurrences of receiver with a cast to
3328       // recv_xtype, since now we know what the type will be.
3329       Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3330       Node* res = _gvn.transform(cast);
3331       if (recv_xtype->is_inlinetypeptr()) {
3332         assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3333         res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3334       }
3335       (*casted_receiver) = res;
3336       assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3337       // (User must make the replace_in_map call.)
3338     }
3339   }
3340 
3341   return fail;
3342 }
3343 
3344 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3345                            float prob) {
3346   Node* want_klass = makecon(tklass);
3347   Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3348   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3349   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3350   set_control(_gvn.transform(new IfTrueNode (iff)));
3351   Node* fail = _gvn.transform(new IfFalseNode(iff));
3352   return fail;
3353 }
3354 
3355 //------------------------------subtype_check_receiver-------------------------
3356 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3357                                        Node** casted_receiver) {
3358   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3359   Node* want_klass = makecon(tklass);
3360 
3361   Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3362 
3363   // Ignore interface type information until interface types are properly tracked.
3364   if (!stopped() && !klass->is_interface()) {
3365     const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3366     const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3367     if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3368       Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3369       if (recv_type->is_inlinetypeptr()) {
3370         cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3371       }
3372       (*casted_receiver) = cast;
3373     }
3374   }
3375 
3376   return slow_ctl;
3377 }
3378 
3379 //------------------------------seems_never_null-------------------------------
3380 // Use null_seen information if it is available from the profile.
3381 // If we see an unexpected null at a type check we record it and force a
3382 // recompile; the offending check will be recompiled to handle nulls.
3383 // If we see several offending BCIs, then all checks in the
3384 // method will be recompiled.
3385 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3386   speculating = !_gvn.type(obj)->speculative_maybe_null();
3387   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3388   if (UncommonNullCast               // Cutout for this technique
3389       && obj != null()               // And not the -Xcomp stupid case?
3390       && !too_many_traps(reason)
3391       ) {
3392     if (speculating) {

3461 
3462 //------------------------maybe_cast_profiled_receiver-------------------------
3463 // If the profile has seen exactly one type, narrow to exactly that type.
3464 // Subsequent type checks will always fold up.
3465 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3466                                              const TypeKlassPtr* require_klass,
3467                                              ciKlass* spec_klass,
3468                                              bool safe_for_replace) {
3469   if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3470 
3471   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3472 
3473   // Make sure we haven't already deoptimized from this tactic.
3474   if (too_many_traps_or_recompiles(reason))
3475     return nullptr;
3476 
3477   // (No, this isn't a call, but it's enough like a virtual call
3478   // to use the same ciMethod accessor to get the profile info...)
3479   // If we have a speculative type use it instead of profiling (which
3480   // may not help us)
3481   ciKlass* exact_kls = spec_klass;
3482   if (exact_kls == nullptr) {
3483     if (java_bc() == Bytecodes::_aastore) {
3484       ciKlass* array_type = nullptr;
3485       ciKlass* element_type = nullptr;
3486       ProfilePtrKind element_ptr = ProfileMaybeNull;
3487       bool flat_array = true;
3488       bool null_free_array = true;
3489       method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3490       exact_kls = element_type;
3491     } else {
3492       exact_kls = profile_has_unique_klass();
3493     }
3494   }
3495   if (exact_kls != nullptr) {// no cast failures here
3496     if (require_klass == nullptr ||
3497         C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3498       // If we narrow the type to match what the type profile sees or
3499       // the speculative type, we can then remove the rest of the
3500       // cast.
3501       // This is a win, even if the exact_kls is very specific,
3502       // because downstream operations, such as method calls,
3503       // will often benefit from the sharper type.
3504       Node* exact_obj = not_null_obj; // will get updated in place...
3505       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
3506                                             &exact_obj);
3507       { PreserveJVMState pjvms(this);
3508         set_control(slow_ctl);
3509         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3510       }
3511       if (safe_for_replace) {
3512         replace_in_map(not_null_obj, exact_obj);
3513       }
3514       return exact_obj;

3604   // If not_null_obj is dead, only null-path is taken
3605   if (stopped()) {              // Doing instance-of on a null?
3606     set_control(null_ctl);
3607     return intcon(0);
3608   }
3609   region->init_req(_null_path, null_ctl);
3610   phi   ->init_req(_null_path, intcon(0)); // Set null path value
3611   if (null_ctl == top()) {
3612     // Do this eagerly, so that pattern matches like is_diamond_phi
3613     // will work even during parsing.
3614     assert(_null_path == PATH_LIMIT-1, "delete last");
3615     region->del_req(_null_path);
3616     phi   ->del_req(_null_path);
3617   }
3618 
3619   // Do we know the type check always succeed?
3620   bool known_statically = false;
3621   if (_gvn.type(superklass)->singleton()) {
3622     const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3623     const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3624     if (subk != nullptr && subk->is_loaded()) {
3625       int static_res = C->static_subtype_check(superk, subk);
3626       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3627     }
3628   }
3629 
3630   if (!known_statically) {
3631     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3632     // We may not have profiling here or it may not help us. If we
3633     // have a speculative type use it to perform an exact cast.
3634     ciKlass* spec_obj_type = obj_type->speculative_type();
3635     if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3636       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3637       if (stopped()) {            // Profile disagrees with this path.
3638         set_control(null_ctl);    // Null is the only remaining possibility.
3639         return intcon(0);
3640       }
3641       if (cast_obj != nullptr) {
3642         not_null_obj = cast_obj;
3643       }
3644     }

3660   record_for_igvn(region);
3661 
3662   // If we know the type check always succeeds then we don't use the
3663   // profiling data at this bytecode. Don't lose it, feed it to the
3664   // type system as a speculative type.
3665   if (safe_for_replace) {
3666     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3667     replace_in_map(obj, casted_obj);
3668   }
3669 
3670   return _gvn.transform(phi);
3671 }
3672 
3673 //-------------------------------gen_checkcast---------------------------------
3674 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3675 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3676 // uncommon-trap paths work.  Adjust stack after this call.
3677 // If failure_control is supplied and not null, it is filled in with
3678 // the control edge for the cast failure.  Otherwise, an appropriate
3679 // uncommon trap or exception is thrown.
3680 // If 'new_cast_failure_map' is supplied and is not null, it is set to a newly cloned map
3681 // when the current map for the success path is updated with information only present
3682 // on the success path and not the cast failure path. The newly cloned map should then be
3683 // used to emit the uncommon trap in the caller.
3684 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node** failure_control, SafePointNode** new_cast_failure_map, bool null_free, bool maybe_larval) {
3685   assert(new_cast_failure_map == nullptr || failure_control != nullptr,
3686          "failure_control must be set when new_failure_map is used");
3687   kill_dead_locals();           // Benefit all the uncommon traps
3688   const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3689   const Type* obj_type = _gvn.type(obj);
3690 
3691   const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3692   const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3693   bool safe_for_replace = (failure_control == nullptr);
3694   assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3695 
3696   // Fast cutout:  Check the case that the cast is vacuously true.
3697   // This detects the common cases where the test will short-circuit
3698   // away completely.  We do this before we perform the null check,
3699   // because if the test is going to turn into zero code, we don't
3700   // want a residual null check left around.  (Causes a slowdown,
3701   // for example, in some objArray manipulations, such as a[i]=a[j].)
3702   if (improved_klass_ptr_type->singleton()) {
3703     const TypeKlassPtr* kptr = nullptr;
3704     if (obj_type->isa_oop_ptr()) {
3705       kptr = obj_type->is_oopptr()->as_klass_type();
3706     } else if (obj->is_InlineType()) {
3707       ciInlineKlass* vk = obj_type->inline_klass();
3708       kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3709     }
3710 
3711     if (kptr != nullptr) {
3712       switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3713       case Compile::SSC_always_true:
3714         // If we know the type check always succeed then we don't use
3715         // the profiling data at this bytecode. Don't lose it, feed it
3716         // to the type system as a speculative type.
3717         obj = record_profiled_receiver_for_speculation(obj);
3718         if (null_free) {
3719           assert(safe_for_replace, "must be");
3720           obj = null_check(obj);
3721         }
3722         return obj;
3723       case Compile::SSC_always_false:
3724         if (null_free) {
3725           assert(safe_for_replace, "must be");
3726           obj = null_check(obj);
3727         }
3728         // It needs a null check because a null will *pass* the cast check.
3729         if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {

3730           bool is_aastore = (java_bc() == Bytecodes::_aastore);
3731           Deoptimization::DeoptReason reason = is_aastore ?
3732             Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3733           builtin_throw(reason);
3734           return top();
3735         } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3736           return null_assert(obj);
3737         }
3738         break; // Fall through to full check
3739       default:
3740         break;
3741       }
3742     }
3743   }
3744 
3745   ciProfileData* data = nullptr;

3746   if (failure_control == nullptr) {        // use MDO in regular case only
3747     assert(java_bc() == Bytecodes::_aastore ||
3748            java_bc() == Bytecodes::_checkcast,
3749            "interpreter profiles type checks only for these BCs");
3750     if (method()->method_data()->is_mature()) {
3751       data = method()->method_data()->bci_to_data(bci());
3752     }
3753   }
3754 
3755   // Make the merge point
3756   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3757   RegionNode* region = new RegionNode(PATH_LIMIT);
3758   Node*       phi    = new PhiNode(region, toop);
3759   _gvn.set_type(region, Type::CONTROL);
3760   _gvn.set_type(phi, toop);
3761 
3762   C->set_has_split_ifs(true); // Has chance for split-if optimization
3763 
3764   // Use null-cast information if it is available
3765   bool speculative_not_null = false;
3766   bool never_see_null = ((failure_control == nullptr)  // regular case only
3767                          && seems_never_null(obj, data, speculative_not_null));
3768 
3769   // Null check; get casted pointer; set region slot 3
3770   Node* null_ctl = top();
3771   Node* not_null_obj = nullptr;
3772   if (null_free) {
3773     assert(safe_for_replace, "must be");
3774     not_null_obj = null_check(obj);
3775   } else {
3776     not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, false /*safe_for_replace*/, speculative_not_null);
3777   }
3778 
3779   // If not_null_obj is dead, only null-path is taken
3780   if (stopped()) {              // Doing instance-of on a null?
3781     set_control(null_ctl);
3782     if (toop->is_inlinetypeptr()) {
3783       return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3784     }
3785     return null();
3786   }
3787   region->init_req(_null_path, null_ctl);
3788   phi   ->init_req(_null_path, null());  // Set null path value
3789   if (null_ctl == top()) {
3790     // Do this eagerly, so that pattern matches like is_diamond_phi
3791     // will work even during parsing.
3792     assert(_null_path == PATH_LIMIT-1, "delete last");
3793     region->del_req(_null_path);
3794     phi   ->del_req(_null_path);
3795   }
3796 
3797   Node* cast_obj = nullptr;
3798   if (improved_klass_ptr_type->klass_is_exact()) {
3799     // The following optimization tries to statically cast the speculative type of the object
3800     // (for example obtained during profiling) to the type of the superklass and then do a
3801     // dynamic check that the type of the object is what we expect. To work correctly
3802     // for checkcast and aastore the type of superklass should be exact.
3803     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3804     // We may not have profiling here or it may not help us. If we have
3805     // a speculative type use it to perform an exact cast.
3806     ciKlass* spec_obj_type = obj_type->speculative_type();
3807     if (spec_obj_type != nullptr || data != nullptr) {
3808       cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, false /*safe_for_replace*/);
3809       if (cast_obj != nullptr) {
3810         if (failure_control != nullptr) // failure is now impossible
3811           (*failure_control) = top();
3812         // adjust the type of the phi to the exact klass:
3813         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3814       }
3815     }
3816   }
3817 
3818   if (cast_obj == nullptr) {
3819     // Generate the subtype check
3820     Node* improved_superklass = superklass;
3821     if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3822       // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3823       // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3824       // Additionally, the benefit would only be minor in non-constant cases.
3825       improved_superklass = makecon(improved_klass_ptr_type);
3826     }
3827     Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);

3828     // Plug in success path into the merge
3829     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3830     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3831     if (failure_control == nullptr) {
3832       if (not_subtype_ctrl != top()) { // If failure is possible
3833         PreserveJVMState pjvms(this);
3834         set_control(not_subtype_ctrl);
3835         bool is_aastore = (java_bc() == Bytecodes::_aastore);
3836         Deoptimization::DeoptReason reason = is_aastore ?
3837           Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3838         builtin_throw(reason);
3839       }
3840     } else {
3841       (*failure_control) = not_subtype_ctrl;
3842     }
3843   }
3844 
3845   region->init_req(_obj_path, control());
3846   phi   ->init_req(_obj_path, cast_obj);
3847 
3848   // Return final merged results
3849   set_control( _gvn.transform(region) );
3850   record_for_igvn(region);
3851 
3852   bool not_inline = !toop->can_be_inline_type();
3853   bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3854   if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3855     // Check if obj has been loaded from an array. Keep obj unchanged for final
3856     // map replacement below.
3857     Node* array_obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3858     Node* array = nullptr;
3859     if (array_obj->isa_Load()) {
3860       Node* address = array_obj->in(MemNode::Address);
3861       if (address->isa_AddP()) {
3862         array = address->as_AddP()->in(AddPNode::Base);
3863       }
3864     } else if (array_obj->is_Phi()) {
3865       Node* region = array_obj->in(0);
3866       // TODO make this more robust (see JDK-8231346)
3867       if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3868         IfNode* iff = region->in(2)->in(0)->isa_If();
3869         if (iff != nullptr) {
3870           iff->is_flat_array_check(&_gvn, &array);
3871         }
3872       }
3873     }
3874     if (array != nullptr) {
3875       const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3876       if (ary_t != nullptr) {
3877         if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3878           // Casting array element to a non-inline-type, mark array as not null-free.
3879           Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3880           if (new_cast_failure_map != nullptr) {
3881             // We want to propagate the improved cast node in the current map. Clone it such that we can still properly
3882             // create the cast failure path in the caller without wrongly making the cast node live there.
3883             *new_cast_failure_map = clone_map();
3884           }
3885           replace_in_map(array, cast);
3886           array = cast;
3887         }
3888         if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3889           // Casting array element to a non-flat-in-array type, mark array as not flat.
3890           Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3891           if (new_cast_failure_map != nullptr && *new_cast_failure_map == nullptr) {
3892             // Same as above.
3893             *new_cast_failure_map = clone_map();
3894           }
3895           replace_in_map(array, cast);
3896           array = cast;
3897         }
3898       }
3899     }
3900   }
3901 
3902   // A merge of null or Casted-NotNull obj
3903   Node* res = _gvn.transform(phi);
3904   if (!stopped() && !res->is_InlineType()) {
3905     res = record_profiled_receiver_for_speculation(res);
3906     if (toop->is_inlinetypeptr() && !maybe_larval) {
3907       Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3908       res = vt;
3909       if (safe_for_replace) {
3910         replace_in_map(obj, vt);
3911         replace_in_map(not_null_obj, vt);
3912         replace_in_map(res, vt);
3913       }
3914     } else if (safe_for_replace) {
3915       replace_in_map(obj, res);
3916     }
3917   }
3918   return res;
3919 }
3920 
3921 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3922   // Load markword
3923   Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3924   Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3925   if (check_lock && !UseCompactObjectHeaders) {
3926     // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3927     // Check if obj is locked
3928     Node* locked_bit = MakeConX(markWord::unlocked_value);
3929     locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3930     Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3931     Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3932     IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3933     _gvn.transform(iff);
3934     Node* locked_region = new RegionNode(3);
3935     Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3936 
3937     // Unlocked: Use bits from mark word
3938     locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3939     mark_phi->init_req(1, mark);
3940 
3941     // Locked: Load prototype header from klass
3942     set_control(_gvn.transform(new IfFalseNode(iff)));
3943     // Make loads control dependent to make sure they are only executed if array is locked
3944     Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3945     Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3946     Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3947     Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3948 
3949     locked_region->init_req(2, control());
3950     mark_phi->init_req(2, proto);
3951     set_control(_gvn.transform(locked_region));
3952     record_for_igvn(locked_region);
3953 
3954     mark = mark_phi;
3955   }
3956 
3957   // Now check if mark word bits are set
3958   Node* mask = MakeConX(mask_val);
3959   Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3960   record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3961   Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3962   return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3963 }
3964 
3965 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3966   return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3967 }
3968 
3969 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3970   // We can't use immutable memory here because the mark word is mutable.
3971   // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3972   // check is moved out of loops (mainly to enable loop unswitching).
3973   Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3974   record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3975   return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3976 }
3977 
3978 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3979   return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3980 }
3981 
3982 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3983   assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3984 
3985   // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3986   if (!vk->has_null_free_non_atomic_layout()) {
3987     return intcon(1); // Always atomic
3988   } else if (!vk->has_null_free_atomic_layout()) {
3989     return intcon(0); // Never atomic
3990   }
3991 
3992   Node* array_klass = load_object_klass(array);
3993   int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3994   Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
3995   Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3996   Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3997   return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3998 }
3999 
4000 Node* GraphKit::atomic_layout_array_test_and_get_layout_kind(Node* array, RegionNode* atomic_region) {
4001   Node* array_klass = load_object_klass(array);
4002   int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
4003   Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
4004   Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
4005   Node* cmp_null_free = _gvn.transform(new CmpINode(layout_kind, intcon(static_cast<jint>(LayoutKind::NULL_FREE_ATOMIC_FLAT))));
4006   Node* bol_null_free = _gvn.transform(new BoolNode(cmp_null_free, BoolTest::eq));
4007   Node* cmp_nullable = _gvn.transform(new CmpINode(layout_kind, intcon(static_cast<jint>(LayoutKind::NULLABLE_ATOMIC_FLAT))));
4008   Node* bol_nullable = _gvn.transform(new BoolNode(cmp_nullable, BoolTest::eq));
4009 
4010   IfNode* iff_null_free = create_and_xform_if(control(), bol_null_free, PROB_FAIR, COUNT_UNKNOWN);
4011   atomic_region->add_req(_gvn.transform(new IfTrueNode(iff_null_free)));
4012   set_control(_gvn.transform(new IfFalseNode(iff_null_free)));
4013 
4014   IfNode* iff_nullable = create_and_xform_if(control(), bol_nullable, PROB_FAIR, COUNT_UNKNOWN);
4015   atomic_region->add_req(_gvn.transform(new IfTrueNode(iff_nullable)));
4016   set_control(_gvn.transform(new IfFalseNode(iff_nullable)));
4017 
4018   return layout_kind;
4019 }
4020 
4021 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
4022 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
4023   RegionNode* region = new RegionNode(3);
4024   Node* null_ctl = top();
4025   null_check_oop(val, &null_ctl);
4026   if (null_ctl != top()) {
4027     PreserveJVMState pjvms(this);
4028     set_control(null_ctl);
4029     {
4030       // Deoptimize if null-free array
4031       BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
4032       inc_sp(nargs);
4033       uncommon_trap(Deoptimization::Reason_null_check,
4034                     Deoptimization::Action_none);
4035     }
4036     region->init_req(1, control());
4037   }
4038   region->init_req(2, control());
4039   set_control(_gvn.transform(region));
4040   record_for_igvn(region);
4041   if (_gvn.type(val) == TypePtr::NULL_PTR) {
4042     // Since we were just successfully storing null, the array can't be null free.
4043     const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
4044     ary_t = ary_t->cast_to_not_null_free();
4045     Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
4046     if (safe_for_replace) {
4047       replace_in_map(ary, cast);
4048     }
4049     ary = cast;
4050   }
4051   return ary;
4052 }
4053 
4054 //------------------------------next_monitor-----------------------------------
4055 // What number should be given to the next monitor?
4056 int GraphKit::next_monitor() {
4057   int current = jvms()->monitor_depth()* C->sync_stack_slots();
4058   int next = current + C->sync_stack_slots();
4059   // Keep the toplevel high water mark current:
4060   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
4061   return current;
4062 }
4063 
4064 //------------------------------insert_mem_bar---------------------------------
4065 // Memory barrier to avoid floating things around
4066 // The membar serves as a pinch point between both control and all memory slices.
4067 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
4068   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
4069   mb->init_req(TypeFunc::Control, control());
4070   mb->init_req(TypeFunc::Memory,  reset_memory());
4071   Node* membar = _gvn.transform(mb);
4072   record_for_igvn(membar);
4073   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));

4174     lock->create_lock_counter(map()->jvms());
4175     increment_counter(lock->counter()->addr());
4176   }
4177 #endif
4178 
4179   return flock;
4180 }
4181 
4182 
4183 //------------------------------shared_unlock----------------------------------
4184 // Emit unlocking code.
4185 void GraphKit::shared_unlock(Node* box, Node* obj) {
4186   // bci is either a monitorenter bc or InvocationEntryBci
4187   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4188   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4189 
4190   if (stopped()) {               // Dead monitor?
4191     map()->pop_monitor();        // Kill monitor from debug info
4192     return;
4193   }
4194   assert(!obj->is_InlineType(), "should not unlock on inline type");
4195 
4196   // Memory barrier to avoid floating things down past the locked region
4197   insert_mem_bar(Op_MemBarReleaseLock);
4198 
4199   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4200   UnlockNode *unlock = new UnlockNode(C, tf);
4201 #ifdef ASSERT
4202   unlock->set_dbg_jvms(sync_jvms());
4203 #endif
4204   uint raw_idx = Compile::AliasIdxRaw;
4205   unlock->init_req( TypeFunc::Control, control() );
4206   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4207   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
4208   unlock->init_req( TypeFunc::FramePtr, frameptr() );
4209   unlock->init_req( TypeFunc::ReturnAdr, top() );
4210 
4211   unlock->init_req(TypeFunc::Parms + 0, obj);
4212   unlock->init_req(TypeFunc::Parms + 1, box);
4213   unlock = _gvn.transform(unlock)->as_Unlock();
4214 
4215   Node* mem = reset_memory();
4216 
4217   // unlock has no side-effects, sets few values
4218   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4219 
4220   // Kill monitor from debug info
4221   map()->pop_monitor( );
4222 }
4223 
4224 //-------------------------------get_layout_helper-----------------------------
4225 // If the given klass is a constant or known to be an array,
4226 // fetch the constant layout helper value into constant_value
4227 // and return null.  Otherwise, load the non-constant
4228 // layout helper value, and return the node which represents it.
4229 // This two-faced routine is useful because allocation sites
4230 // almost always feature constant types.
4231 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4232   const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4233   if (!StressReflectiveCode && klass_t != nullptr) {
4234     bool xklass = klass_t->klass_is_exact();
4235     bool can_be_flat = false;
4236     const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4237     if (UseArrayFlattening && !xklass && ary_type != nullptr) {
4238       // Don't constant fold if the runtime type might be a flat array but the static type is not.
4239       const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4240       can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4241     }
4242     if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4243       jint lhelper;
4244       if (klass_t->is_flat()) {
4245         lhelper = ary_type->flat_layout_helper();
4246       } else if (klass_t->isa_aryklassptr()) {
4247         BasicType elem = ary_type->elem()->array_element_basic_type();
4248         if (is_reference_type(elem, true)) {
4249           elem = T_OBJECT;
4250         }
4251         lhelper = Klass::array_layout_helper(elem);
4252       } else {
4253         lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4254       }
4255       if (lhelper != Klass::_lh_neutral_value) {
4256         constant_value = lhelper;
4257         return (Node*) nullptr;
4258       }
4259     }
4260   }
4261   constant_value = Klass::_lh_neutral_value;  // put in a known value
4262   Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
4263   return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4264 }
4265 
4266 // We just put in an allocate/initialize with a big raw-memory effect.
4267 // Hook selected additional alias categories on the initialization.
4268 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4269                                 MergeMemNode* init_in_merge,
4270                                 Node* init_out_raw) {
4271   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4272   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4273 
4274   Node* prevmem = kit.memory(alias_idx);
4275   init_in_merge->set_memory_at(alias_idx, prevmem);
4276   if (init_out_raw != nullptr) {
4277     kit.set_memory(init_out_raw, alias_idx);
4278   }
4279 }
4280 
4281 //---------------------------set_output_for_allocation-------------------------
4282 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4283                                           const TypeOopPtr* oop_type,
4284                                           bool deoptimize_on_exception) {
4285   int rawidx = Compile::AliasIdxRaw;
4286   alloc->set_req( TypeFunc::FramePtr, frameptr() );
4287   add_safepoint_edges(alloc);
4288   Node* allocx = _gvn.transform(alloc);
4289   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4290   // create memory projection for i_o
4291   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4292   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4293 
4294   // create a memory projection as for the normal control path
4295   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4296   set_memory(malloc, rawidx);
4297 
4298   // a normal slow-call doesn't change i_o, but an allocation does
4299   // we create a separate i_o projection for the normal control path
4300   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4301   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4302 
4303   // put in an initialization barrier
4304   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4305                                                  rawoop)->as_Initialize();
4306   assert(alloc->initialization() == init,  "2-way macro link must work");
4307   assert(init ->allocation()     == alloc, "2-way macro link must work");
4308   {
4309     // Extract memory strands which may participate in the new object's
4310     // initialization, and source them from the new InitializeNode.
4311     // This will allow us to observe initializations when they occur,
4312     // and link them properly (as a group) to the InitializeNode.
4313     assert(init->in(InitializeNode::Memory) == malloc, "");
4314     MergeMemNode* minit_in = MergeMemNode::make(malloc);
4315     init->set_req(InitializeNode::Memory, minit_in);
4316     record_for_igvn(minit_in); // fold it up later, if possible
4317     _gvn.set_type(minit_in, Type::MEMORY);
4318     Node* minit_out = memory(rawidx);
4319     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4320     int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4321     // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4322     // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4323     // multiple projections as a result.
4324     set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4325     int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4326     set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4327     if (oop_type->isa_aryptr()) {
4328       // Initially all flat array accesses share a single slice
4329       // but that changes after parsing. Prepare the memory graph so
4330       // it can optimize flat array accesses properly once they
4331       // don't share a single slice.
4332       assert(C->flat_accesses_share_alias(), "should be set at parse time");
4333       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4334       int            elemidx  = C->get_alias_index(telemref);
4335       const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4336       if (alias_adr_type->is_flat()) {
4337         C->set_flat_accesses();
4338       }
4339       hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4340     } else if (oop_type->isa_instptr()) {
4341       ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4342       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4343         ciField* field = ik->nonstatic_field_at(i);
4344         if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4345           continue;  // do not bother to track really large numbers of fields
4346         // Find (or create) the alias category for this field:
4347         int fieldidx = C->alias_type(field)->index();
4348         hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4349       }
4350     }
4351   }
4352 
4353   // Cast raw oop to the real thing...
4354   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4355   javaoop = _gvn.transform(javaoop);
4356   C->set_recent_alloc(control(), javaoop);
4357   assert(just_allocated_object(control()) == javaoop, "just allocated");
4358 
4359 #ifdef ASSERT

4371       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4372     }
4373   }
4374 #endif //ASSERT
4375 
4376   return javaoop;
4377 }
4378 
4379 //---------------------------new_instance--------------------------------------
4380 // This routine takes a klass_node which may be constant (for a static type)
4381 // or may be non-constant (for reflective code).  It will work equally well
4382 // for either, and the graph will fold nicely if the optimizer later reduces
4383 // the type to a constant.
4384 // The optional arguments are for specialized use by intrinsics:
4385 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4386 //  - If 'return_size_val', report the total object size to the caller.
4387 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4388 Node* GraphKit::new_instance(Node* klass_node,
4389                              Node* extra_slow_test,
4390                              Node* *return_size_val,
4391                              bool deoptimize_on_exception,
4392                              InlineTypeNode* inline_type_node) {
4393   // Compute size in doublewords
4394   // The size is always an integral number of doublewords, represented
4395   // as a positive bytewise size stored in the klass's layout_helper.
4396   // The layout_helper also encodes (in a low bit) the need for a slow path.
4397   jint  layout_con = Klass::_lh_neutral_value;
4398   Node* layout_val = get_layout_helper(klass_node, layout_con);
4399   bool  layout_is_con = (layout_val == nullptr);
4400 
4401   if (extra_slow_test == nullptr)  extra_slow_test = intcon(0);
4402   // Generate the initial go-slow test.  It's either ALWAYS (return a
4403   // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4404   // case) a computed value derived from the layout_helper.
4405   Node* initial_slow_test = nullptr;
4406   if (layout_is_con) {
4407     assert(!StressReflectiveCode, "stress mode does not use these paths");
4408     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4409     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4410   } else {   // reflective case
4411     // This reflective path is used by Unsafe.allocateInstance.
4412     // (It may be stress-tested by specifying StressReflectiveCode.)
4413     // Basically, we want to get into the VM is there's an illegal argument.
4414     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4415     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4416     if (extra_slow_test != intcon(0)) {
4417       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4418     }
4419     // (Macro-expander will further convert this to a Bool, if necessary.)

4430 
4431     // Clear the low bits to extract layout_helper_size_in_bytes:
4432     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4433     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4434     size = _gvn.transform( new AndXNode(size, mask) );
4435   }
4436   if (return_size_val != nullptr) {
4437     (*return_size_val) = size;
4438   }
4439 
4440   // This is a precise notnull oop of the klass.
4441   // (Actually, it need not be precise if this is a reflective allocation.)
4442   // It's what we cast the result to.
4443   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4444   if (!tklass)  tklass = TypeInstKlassPtr::OBJECT;
4445   const TypeOopPtr* oop_type = tklass->as_instance_type();
4446 
4447   // Now generate allocation code
4448 
4449   // The entire memory state is needed for slow path of the allocation
4450   // since GC and deoptimization can happen.
4451   Node *mem = reset_memory();
4452   set_all_memory(mem); // Create new memory state
4453 
4454   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4455                                          control(), mem, i_o(),
4456                                          size, klass_node,
4457                                          initial_slow_test, inline_type_node);
4458 
4459   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4460 }
4461 
4462 //-------------------------------new_array-------------------------------------
4463 // helper for newarray and anewarray
4464 // The 'length' parameter is (obviously) the length of the array.
4465 // The optional arguments are for specialized use by intrinsics:
4466 //  - If 'return_size_val', report the non-padded array size (sum of header size
4467 //    and array body) to the caller.
4468 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4469 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
4470                           Node* length,         // number of array elements
4471                           int   nargs,          // number of arguments to push back for uncommon trap
4472                           Node* *return_size_val,
4473                           bool deoptimize_on_exception,
4474                           Node* init_val) {
4475   jint  layout_con = Klass::_lh_neutral_value;
4476   Node* layout_val = get_layout_helper(klass_node, layout_con);
4477   bool  layout_is_con = (layout_val == nullptr);
4478 
4479   if (!layout_is_con && !StressReflectiveCode &&
4480       !too_many_traps(Deoptimization::Reason_class_check)) {
4481     // This is a reflective array creation site.
4482     // Optimistically assume that it is a subtype of Object[],
4483     // so that we can fold up all the address arithmetic.
4484     layout_con = Klass::array_layout_helper(T_OBJECT);
4485     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4486     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4487     { BuildCutout unless(this, bol_lh, PROB_MAX);
4488       inc_sp(nargs);
4489       uncommon_trap(Deoptimization::Reason_class_check,
4490                     Deoptimization::Action_maybe_recompile);
4491     }
4492     layout_val = nullptr;
4493     layout_is_con = true;
4494   }
4495 
4496   // Generate the initial go-slow test.  Make sure we do not overflow
4497   // if length is huge (near 2Gig) or negative!  We do not need
4498   // exact double-words here, just a close approximation of needed
4499   // double-words.  We can't add any offset or rounding bits, lest we
4500   // take a size -1 of bytes and make it positive.  Use an unsigned
4501   // compare, so negative sizes look hugely positive.
4502   int fast_size_limit = FastAllocateSizeLimit;
4503   if (layout_is_con) {
4504     assert(!StressReflectiveCode, "stress mode does not use these paths");
4505     // Increase the size limit if we have exact knowledge of array type.
4506     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4507     fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);


4508   }
4509 
4510   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4511   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4512 
4513   // --- Size Computation ---
4514   // array_size = round_to_heap(array_header + (length << elem_shift));
4515   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4516   // and align_to(x, y) == ((x + y-1) & ~(y-1))
4517   // The rounding mask is strength-reduced, if possible.
4518   int round_mask = MinObjAlignmentInBytes - 1;
4519   Node* header_size = nullptr;
4520   // (T_BYTE has the weakest alignment and size restrictions...)
4521   if (layout_is_con) {
4522     int       hsize  = Klass::layout_helper_header_size(layout_con);
4523     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
4524     bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4525     if ((round_mask & ~right_n_bits(eshift)) == 0)
4526       round_mask = 0;  // strength-reduce it if it goes away completely
4527     assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4528     int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4529     assert(header_size_min <= hsize, "generic minimum is smallest");
4530     header_size = intcon(hsize);
4531   } else {
4532     Node* hss   = intcon(Klass::_lh_header_size_shift);
4533     Node* hsm   = intcon(Klass::_lh_header_size_mask);
4534     header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4535     header_size = _gvn.transform(new AndINode(header_size, hsm));
4536   }
4537 
4538   Node* elem_shift = nullptr;
4539   if (layout_is_con) {
4540     int eshift = Klass::layout_helper_log2_element_size(layout_con);
4541     if (eshift != 0)
4542       elem_shift = intcon(eshift);
4543   } else {
4544     // There is no need to mask or shift this value.
4545     // The semantics of LShiftINode include an implicit mask to 0x1F.
4546     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4547     elem_shift = layout_val;

4596   }
4597   Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4598 
4599   if (return_size_val != nullptr) {
4600     // This is the size
4601     (*return_size_val) = non_rounded_size;
4602   }
4603 
4604   Node* size = non_rounded_size;
4605   if (round_mask != 0) {
4606     Node* mask1 = MakeConX(round_mask);
4607     size = _gvn.transform(new AddXNode(size, mask1));
4608     Node* mask2 = MakeConX(~round_mask);
4609     size = _gvn.transform(new AndXNode(size, mask2));
4610   }
4611   // else if round_mask == 0, the size computation is self-rounding
4612 
4613   // Now generate allocation code
4614 
4615   // The entire memory state is needed for slow path of the allocation
4616   // since GC and deoptimization can happen.
4617   Node *mem = reset_memory();
4618   set_all_memory(mem); // Create new memory state
4619 
4620   if (initial_slow_test->is_Bool()) {
4621     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4622     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4623   }
4624 
4625   const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4626   const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4627 
4628   Node* raw_init_value = nullptr;
4629   if (init_val != nullptr) {
4630     // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4631     if (ary_type->is_flat()) {
4632       initial_slow_test = intcon(1);
4633     }
4634 
4635     if (UseCompressedOops) {
4636       // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4637       init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4638       Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4639       Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4640       raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4641     } else {
4642       raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4643     }
4644   }
4645 
4646   Node* valid_length_test = _gvn.intcon(1);
4647   if (ary_type->isa_aryptr()) {
4648     BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4649     jint max = TypeAryPtr::max_array_length(bt);
4650     Node* valid_length_cmp  = _gvn.transform(new CmpUNode(length, intcon(max)));
4651     valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4652   }
4653 
4654   // Create the AllocateArrayNode and its result projections
4655   AllocateArrayNode* alloc
4656     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4657                             control(), mem, i_o(),
4658                             size, klass_node,
4659                             initial_slow_test,
4660                             length, valid_length_test,
4661                             init_val, raw_init_value);
4662   // Cast to correct type.  Note that the klass_node may be constant or not,
4663   // and in the latter case the actual array type will be inexact also.
4664   // (This happens via a non-constant argument to inline_native_newArray.)
4665   // In any case, the value of klass_node provides the desired array type.
4666   const TypeInt* length_type = _gvn.find_int_type(length);
4667   if (ary_type->isa_aryptr() && length_type != nullptr) {
4668     // Try to get a better type than POS for the size
4669     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4670   }
4671 
4672   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4673 
4674   array_ideal_length(alloc, ary_type, true);
4675   return javaoop;
4676 }
4677 
4678 // The following "Ideal_foo" functions are placed here because they recognize
4679 // the graph shapes created by the functions immediately above.
4680 
4681 //---------------------------Ideal_allocation----------------------------------

4776 void GraphKit::add_parse_predicates(int nargs) {
4777   if (ShortRunningLongLoop) {
4778     // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4779     // walking up from the loop.
4780     add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4781   }
4782   if (UseLoopPredicate) {
4783     add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4784     if (UseProfiledLoopPredicate) {
4785       add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4786     }
4787   }
4788   if (UseAutoVectorizationPredicate) {
4789     add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4790   }
4791   // Loop Limit Check Predicate should be near the loop.
4792   add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4793 }
4794 
4795 void GraphKit::sync_kit(IdealKit& ideal) {
4796   reset_memory();
4797   set_all_memory(ideal.merged_memory());
4798   set_i_o(ideal.i_o());
4799   set_control(ideal.ctrl());
4800 }
4801 
4802 void GraphKit::final_sync(IdealKit& ideal) {
4803   // Final sync IdealKit and graphKit.
4804   sync_kit(ideal);
4805 }
4806 
4807 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4808   Node* len = load_array_length(load_String_value(str, set_ctrl));
4809   Node* coder = load_String_coder(str, set_ctrl);
4810   // Divide length by 2 if coder is UTF16
4811   return _gvn.transform(new RShiftINode(len, coder));
4812 }
4813 
4814 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4815   int value_offset = java_lang_String::value_offset();
4816   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4817                                                      false, nullptr, Type::Offset(0));
4818   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4819   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::BotPTR,
4820                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4821                                                   ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4822   Node* p = basic_plus_adr(str, str, value_offset);
4823   Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4824                               IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4825   return must_be_not_null(load, true);
4826 }
4827 
4828 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4829   if (!CompactStrings) {
4830     return intcon(java_lang_String::CODER_UTF16);
4831   }
4832   int coder_offset = java_lang_String::coder_offset();
4833   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4834                                                      false, nullptr, Type::Offset(0));
4835   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4836 
4837   Node* p = basic_plus_adr(str, str, coder_offset);
4838   Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4839                               IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4840   return load;
4841 }
4842 
4843 void GraphKit::store_String_value(Node* str, Node* value) {
4844   int value_offset = java_lang_String::value_offset();
4845   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4846                                                      false, nullptr, Type::Offset(0));
4847   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4848 
4849   access_store_at(str,  basic_plus_adr(str, value_offset), value_field_type,
4850                   value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4851 }
4852 
4853 void GraphKit::store_String_coder(Node* str, Node* value) {
4854   int coder_offset = java_lang_String::coder_offset();
4855   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4856                                                      false, nullptr, Type::Offset(0));
4857   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4858 
4859   access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4860                   value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4861 }
4862 
4863 // If input and output memory types differ, capture the whole memory to preserve
4864 // the dependency between preceding and subsequent loads/stores.
4865 // For example, the following program:
4866 //  StoreB
4867 //  compress_string
4868 //  LoadB
4869 // has this memory graph (use->def):
4870 //  LoadB -> compress_string -> CharMem
4871 //             ... -> StoreB -> ByteMem
4872 // The intrinsic hides the dependency between LoadB and StoreB, causing
4873 // the load to read from memory not containing the result of the StoreB.
4874 // The correct memory graph should look like this:
4875 //  LoadB -> compress_string -> MergeMem -> StoreB
4876 Node* GraphKit::capture_memory(const TypePtr*& combined_type, const TypePtr* src_type, const TypePtr* dst_type) {

4979   i_char->init_req(2, AddI(i_char, intcon(2)));
4980 
4981   set_control(IfFalse(iff));
4982   set_memory(st, TypeAryPtr::BYTES);
4983 }
4984 
4985 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4986   if (!field->is_constant()) {
4987     return nullptr; // Field not marked as constant.
4988   }
4989   ciInstance* holder = nullptr;
4990   if (!field->is_static()) {
4991     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4992     if (const_oop != nullptr && const_oop->is_instance()) {
4993       holder = const_oop->as_instance();
4994     }
4995   }
4996   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4997                                                         /*is_unsigned_load=*/false);
4998   if (con_type != nullptr) {
4999     Node* con = makecon(con_type);
5000     if (field->type()->is_inlinetype()) {
5001       con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
5002     } else if (con_type->is_inlinetypeptr()) {
5003       con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
5004     }
5005     return con;
5006   }
5007   return nullptr;
5008 }
5009 
5010 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
5011   const Type* obj_type = obj->bottom_type();
5012   const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
5013   if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
5014     const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
5015     Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
5016     obj = casted_obj;
5017   }
5018   if (!maybe_larval && sig_type->is_inlinetypeptr()) {
5019     obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
5020   }
5021   return obj;
5022 }
< prev index next >